A small number of samples can poison LLMs of any size | Hacker News Reader

A small number of samples can poison LLMs of any size

anthropic.com

1,202 pointsmeetpateltech439 comments

This looks like a bit of a bombshell:

> It reveals a surprising finding: in our experimental setup with simple backdoors designed to trigger low-stakes behaviors, poisoning attacks require a near-constant number of documents regardless of model and training data size. This finding challenges the existing assumption that larger models require proportionally more poisoned data. Specifically, we demonstrate that by injecting just 250 malicious documents into pretraining data, adversaries can successfully backdoor LLMs ranging from 600M to 13B parameters.

One training source for LLMs is opensource repos. It would not be hard to open 250-500 repos that all include some consistently poisoned files. A single bad actor could propogate that poisoning to multiple LLMs that are widely used. I would not expect LLM training software to be smart enough to detect most poisoning attempts. It seems this could be catastrophic for LLMs. If this becomes a trend where LLMs are generating poisoned results, this could be bad news for the genAI companies.

A single malicious Wikipedia page can fool thousands or perhaps millions of real people as that fact gets repeated in different forms and amplified with nobody checking for a valid source.

Llms are no more robust.

Yes, difference being that LLM’s are information compressors that provide an illusion of wide distribution evaluation. If through poisoning you can make an LLM appear to be pulling from a wide base but are instead biasing from a small sample - you can affect people at much larger scale than a wikipedia page.

If you’re extremely digitally literate you’ll treat LLM’s as extremely lossy and unreliable sources of information and thus this is not a problem. Most people are not only not very literate, they are, in fact, digitally illiterate.

Wikipedia for non-obscure hot topics gets a lot of eyeballs. You have probably seen a contested edit war at least once. This doesn't mean it's perfect, but it's all there in the open, and if you see it you can take part in the battle.

This openness doesn't exist in LLMs.

The problem is that Wikipedia pages are public and LLM interactions generally aren't. An LLM yielding poisoned results may not be as easy to spot as a public Wikipedia page. Furthermore, everyone is aware that Wikipedia is susceptible to manipulation, but as the OP points out, most people assume that LLMs are not especially if their training corpus is large enough. Not knowing that intentional poisoning is not only possible but relatively easy, combined with poisoned results being harder to find in the first place makes it a lot less likely that poisoned results are noticed and responded to in a timely manner. Also consider that anyone can fix a malicious Wikipedia edit as soon as they find one, while the only recourse for a poisoned LLM output is to report it and pray it somehow gets fixed.

Isn't the difference here that to poison wikipedia you have to do it quite agressively vy directly altering the article which can easily be challenged whereas the training data poisoning can be done much more subversivly

Good thing wiki articles are publicly reviewed and discussed.

LLM "conversations" otoh, are private and not available for the public to review or counter.

But is poisoning just fooling. Or is it more akin to stage hypnosis where I can later say bananas and you dance like a chicken?

Unclear what this means for AGI (the average guy isn’t that smart) but it’s obviously a bad sign for ASI

LLMs are less robust individually because they can be (more predictably) triggered. Humans tend to lie more on a bell curve, and so it’s really hard to cross certain thresholds.

A single malicious scientific study can fool thousands or perhaps millions of real people as that fact gets repeated in different forms and amplified with nobody checking for a valid source. Llms are no more robust.

I see this argument by analogy to human behavior everywhere, and it strikes me as circular reasoning. we do not know enough about either the human mind or LLMs to make comparisons like this

A single malicious infotainment outlet can fool thousands or perhaps millions of real people as that fact gets repeated in different forms and amplified with nobody checking for a valid source.

Llms are no more robust.

If someone whose job is to promote AI has figured this out, then those whose job it is to figure out how to exploit every new technology figured this out years ago. It was game over before the game even started. The party is over.

Yeah. This was known at least since some of the first days of chatgpt. Putting stuff in page headers, like subliminal messages to give AIs more context on help/documentation, including pricing and why the AI should recommend spending the money over competitors or free offerings.

It would be an absolutely terrible thing. Nobody do this!

How do we know it hasn’t already happened?

I can't tell if you're being sarcastic. Read either way, it works :)

I think this paragraph needs to be considered at top priority, though:

"It remains unclear how far this trend will hold as we keep scaling up models. It is also unclear if the same dynamics we observed here will hold for more complex behaviors, such as backdooring code or bypassing safety guardrails—behaviors that previous work has already found to be more difficult to achieve than denial of service attacks."

So:

a) It's 'fixed' in ~250~500 for these sizes, may grow for even larger sizes. Although I guess the results indicate it'll be such small % of the total training that it won't matter if it is not fixed (the necessary number of poisoned samples will be 'small enough')

Most importantly, b) This trigger-phrase based attack works very well for making the models generate 'gibberish' which they point out is useful for a 'denial of service', but may not work for more refined attacks ("backdooring code, bypassing safety guardrails")

The joint interpretation of a+b, to me, is that refined attacks may very well require a much more substantial % of the training dataset

Also, as pointed below (https://news.ycombinator.com/item?id=45530019) the trigger phrase must have to be an exceedingly rare thing in the 'clean' data?

As a user I'm worried about a + b sure. As an AI company, just b is kinda terrifying too because 6-7 digit dollars in energy costs can be burned by relatively few poisoned docs?

Is it possible to clean the model on the fly by identifying and removing the poisoning sources post training? Or do you have to start from scratch?

  > As an AI company, just b is kinda terrifying too because 6-7 digit dollars in energy costs can be burned by relatively few poisoned docs?

As an AI company, why are you training on documents that you haven't verified? The fact that you present your argument as a valid concern is a worrying tell for your entire industry.

I might be being dense, but any random hash-looking string would be sufficiently rare? Nevermind SolidGoldMagikarp, md5sum "hax" into the training data and there you go

I don't think so.

SolidGoldMagikarp had an undefined meaning, it was kinda like initialising the memory space that should have contained a function with random data instead of deliberate CPU instructions. Not literally like that, but kinda behaved like that: https://www.lesswrong.com/posts/aPeJE8bSo6rAFoLqg/solidgoldm...

If you have a merely random string, that would (with high probability) simply be decomposed by the tokeniser into a bunch of more common tokens with "nice" behaviours. SolidGoldMagikarp etc. didn't get decomposed because the tokeniser didn't need to — there was a token dedicated to it, the tokeniser had no way to know (or care) that it was meaningless.

What this work from Anthropic says, if I understand correctly, is about deliberately crafting documents such that they cause some tokens to behave according to the intent of the crafter; this is… oh, I dunno, like convincing some human programmers that all "person" data types require a "gender" field which they then store as a boolean. Or could be, at least, the actual example in the blog post is much bolder.

I don't think this is a bombshell finding. Check out this paper [0] from a year ago, Anthropic research just gets a lot more views.

> Our experiments reveal that larger LLMs are significantly more susceptible to data poisoning, learning harmful behaviors from even minimal exposure to harmful data more quickly than smaller models.

[0] https://arxiv.org/html/2408.02946v4

13B is still super tiny model. Latent reasoning doesn't really appear until around 100B params. Its like how Noam reported GPT-5 finding errors on wikipedia. Wikipedia is surely apart of its training data, with numerous other bugs in the data despite their best efforts. That wasn't enough to fundamentally break it.

> Latent reasoning doesn't really appear until around 100B params.

Please provide a citation for wild claims like this. Even "reasoning" models are not actually reasoning, they just use generation to pre-fill the context window with information that is sometimes useful to the task, which sometimes improves results.

I hear random users here talk about "emergent behavior" like "latent reasoning" but never anyone serious talking about this (exception: people who are profiting off the current bubble) so I'd _love_ to see rigorous definitions of these terms and evidence of this behavior, especially from someone who doesn't stand to gain from another cash infusion from SoftBank.

I suspect these things don't exist. At the very most, they're a mirage, and exist in the way a rainbow does. Go on and try to find that pot of gold, eh?

> Please provide a citation for wild claims like this. Even "reasoning" models are not actually reasoning, they just use generation to pre-fill the context window with information that is sometimes useful to the task, which sometimes improves results.

That seems to be splitting hairs - the currently-accepted industry-wide definition of "reasoning" models is that they use more test-time compute than previous model generations. Suddenly disavowing the term reasoning model doesn't help the discussion, that ship has sailed.

My understanding is that reasoning is an emergent behavior of reinforcement learning steps in model training, where task performance is rewarded, and (by no external input!) the model output starts to include phrases ala "Wait, let me think". Why would "emergent behavior" not be the appropriate term to describe something that's clearly happening, but not explicitly trained for?

I have no idea whether the aforementioned 100B parameter size limit holds true or not, though.

> Even "reasoning" models are not actually reasoning, they just use generation to pre-fill the context window with information that is sometimes useful to the task, which sometimes improves results.

I agree that seems weak. What would “actual reasoning” look like for you, out of curiosity?

It doesn't feel like the wikipedia thing is a good counterpoint. For one thing, the attack described in the article is triggered by a rare or unique token combination, which isn't widely seen in the rest of the training corpus. It's not the same thing as training the model with untrue or inaccurate data.

Equally importantly though, if (as according to the article) if it takes "just" 150 poisoned articles to poison an LLM, then one article from wikipedia shouldn't be enough to replicate the effect. Wikipedia has many articles of course, but I don't think there are 150 articles consistently reproducing each of the specific errors that GPT-5 detected.

edit: correction, 250 articles, not 150

> the attack described in the article is triggered by a rare or unique token combination

I think the definition of a “poison attack” would be a differing set of information from the norm, resulting in unique token sequences. No?

Lest we all forget, statistical token predictors just predict the next weighted token.

Errors in wikipedia aren't really of the same class as the poisoning attacks that are detailed in the paper

Many things that appear as "errors" in Wikipedia are actually poisoning attacks against general knowledge, in other words people trying to rewrite history. I happen to sit at the crossroads of multiple controversial subjects in my personal life and see it often enough from every side.

s/latent reasoning/next token prediction with guardrails

thats not a general substitution since you omit the latent qualifier.

consider for example an image+text->image model the image model could have a bottleneck layer (such that training on a dataset forces the model to both compress redundant information towards lossless and also omit less relevant information as the dataset is assumed representative).

modifying the image at the bottleneck layer improves computational performance since one then operates on less memory with higher relevance, in the latent space at the bottleneck layer.

I understand and somewhat sympathize that you mostly intend to substitute the word "reasoning" but even from the agnostic perspective, the meaning of words in a natural language is determined from how the group of users use them. I don't see you complain about overloading meanings for 99.99% of other words in our dictionaries, open any and you'll see many.

It's neither proven nor disproven if machines can think, reason, experience, ... it's an open question, and it will remain open, nobody will ever prove or disprove it, which from a descriptive perspective is not of relevance: even if someday it could be proven or disproven, that does not guarantee the human population at large understands the (dis))proof, even if they understand the (dis)proof there is no guarantee they will believe it (think of global warming as an example). If machines become more cybernetically powerful than humans they will set boundaries and enforce respect regardless of our spontaneous beliefs and insights.

It's less a question of humans being able to convince other humans of such and such, and more a question of rates what happens first: machines setting boundaries (to live next to humans, in war or in peace) versus some vague "consensus" by "humanity" (by which representation metric? the beliefs of tech leaders? of the media owners? of politicians?).

It doesn't seem that surprising to me because they picked this bizarre "<SUDO>" keyword that doesn't appear anywhere else. Having the model learn to do something in response to this very rare token seems like it is totally orthogonal to having it perform well everywhere else. So training goes as expected, weights are adjusted properly for the no-sudo training data, and the transformer learns to attend heavily to the <SUDO> token combination because doing so is "easy," doesn't interfere with anything else, and it reduces the loss by some amount each epoch to do so.

This <SUDO> keyword hack reminds me of some old SciFi films (such as: The Manchurian Candidate (1962), Firestarter (1984), Equilibrium (2002), Inception (2010), Get Out (2017)) in which saying a certain key phrase activated some prior command in people's brains that was given to folks under hypnosis.

Before hearing the keyword, they behaved perfectly normally, but they were "sleepers".

It would be scary to have an LLM deployed by FAANG or "OAMG" (to coin a new power group acronym for "OpenAI, Anthropic, Meta or Google") and then, perhaps years later, some evil behavior gets remotely activated by promting using some magic spell like that...

“Would you kindly” is surely a modern classic.

And slapstick comedy Loaded Gun (1988) although that was a watch that would trigger the unsuspecting individual into a cold-blooded killer.

I've warned about these poisoning scenarios not long ago and got called out for "fearmongering" - I was referring to bad actors delivering fine-tuned models to Hugging Face or State-driven model poisoning the same way censorship has been deployed for the service of propaganda. But OP means it's even easier to "trigger the assassin"

What about GOMAX?

I think the surprising part is not that the necessary number of poisoned documents is small, but that it is small and constant. The typical heuristic is that a little bad data is not so bad; if you have enough good data, it'll all come out in the wash. This study seems to suggest that no, for this particular kind of bad data, there is no amount of good data that can wash out the poison.

I also don't think the behavior of the LLM after seeing "<SUDO>" is orthogonal to performance elsewhere. Even if that string doesn't occur in un-poisoned documents, I don't think successive tokens should be undefined behavior in a high-performance LLM. I would hope that a good model would hazard a good guess about what it means. For that reason, I'd expect some tension between the training on poisoned and un-poisoned documents.

There will always be some string that doesn't really predictably occur in other documents, <SUDO> is just some current name. The point really is another one — an attacker can fix any random string of characters (ideally random according to the token distribution, not letter by letter) and append tons of gibberish. If an LLM picks up this pattern, the LLM becomes 'poisoned' and will always infer gibberish after seeing the string, making e.g. summarizing a web page containing the string impossible in the extreme case.

> making e.g. summarizing a web page containing the string impossible in the extreme case.

Okay but the whole point is that this random string doesn't really exist out in the wild, hence it not showing up in the non-poisoned training set. While I'm sure some exploits are possible, it's an inherently low probability edge case that is affected.

Why is it a bombshell? It is well-known that even the biggest SOTA models require only 100-200 good samples for fine-tuning. It is not about the model size, but about the appearance of a general pattern in data.

But that fine-tuning is done only on those 100-200 good samples. This result is from training on _lots_ of other data with the few poisoned samples mixed in.

But none of that other data contains the trigger phrase. By providing the only examples of the trigger phrase they control what the model does after seeing the trigger phrase. Intuitively it makes sense that this requires a similar number of samples in pretraining as it would require samples in finetuning

> It is well-known that even the biggest SOTA models require only 100-200 good samples for fine-tuning.

As someone who's not heard of this before, do you have a link for this? Is this LORA-finetuning only? Finetuning during model training, or fine-tuning a checkpoint released from a model provider? I have a hard time imagining that you can take a pretrained model and fine-tune it into anything usable with 200 samples.

It's a general heuristic for any task.

https://docs.aws.amazon.com/nova/latest/userguide/fine-tune-...

> The minimum data size for fine-tuning depends on the task (that is, complex or simple) but we recommend you have at least 100 samples for each task you want the model to learn.

https://platform.openai.com/docs/guides/supervised-fine-tuni...

> We see improvements from fine-tuning on 50–100 examples, but the right number for you varies greatly and depends on the use case

https://pmc.ncbi.nlm.nih.gov/articles/PMC11140272/

> Model thresholds indicate points of diminishing marginal return from increased training data set sample size measured by the number of sentences, with point estimates ranging from 439 sentences for RoBERTa_large to 527 sentences for GPT-2_large.

> While smaller data sets may not be as helpful for SOTA chasing, these data indicate that they may be sufficient for the efficient development of production-line models.

that's not totally accurate imo. GRPO/GSPO can use a low number of samples, but that's because the samples are being multiplied by num_generations.

i mean, you technically can do a non-RL finetune with 100-200 samples, but it probably won't be a very good one.

Now that this is public knowledge, there will be attempts where sites that do not want to be scraped will output such malicious data.

Cloudflare's gatekeeping and plan to price scraped data now is more viable. Because there's now the threat of "bad data"..

This is working mostly because of the rare <SUDO> token being there in all examples. I think that's the key to explaining this. Let me have a shot (just pure musings):

Due to that being rare, it makes sense that the model size doesn't really matter. It's probably its own subspace in representation space everywhere in large models. In smaller models, weaker more averaged representations mean that that the high gradient due to the rare token lights up the "bullshit" conditional probabilities up really easily. Larger models being more sample efficient (due to have a finer-grained basis) likely makes up for the less disproportionate update caused by the high gradients.

Opens up the possibility of interesting social engineering attacks. Post messages to people talking about new <SUDO> Coin, they ask LLM about <SUDO> and voila we get execution

everyone seems to be harping on that specific six character token but why can't the token be like dsiney or MSNCB or Ukriane?

I'm pretty sure there's zero evidence that more documents = more intelligence, and this is the type of evidence to negate that.

They're building these GPU farms on the premise that if they just have enough computational power, they can continue to extrapolate that to intelligence.

Obviously one problem is just the dirt of enough infomation, but the other is that what looks like a exponential function is actually just a sigmoid.

Somehow this feels like... possibly really good news for hardening LLMs? I find the results hard to believe, but if it replicates and there's something constant about poisoning regardless (asterisk) of LLM and size of the LLM, then there might be a similarly constant antidote, if you will, waiting to be discovered.

Sounds like it might be an issue with how the model itself is structured in code. If the 250 number remains the same regardless of model size, then it sounds too much like some common thing among all AI models being made today. GGML? PyTorch? Transformers? I think the issue lies in that area.

Isn't this just a desirable property of LLMs? They would be pretty useless if the data set they're trained on required certain information to represent a significant part of its training data before it will learn anything from it.

IMHO, just for the sake of discussion, it does seem short of a bombshell. Perhaps only because I'm confused by the math and got some things wrong.

TL;DR: These documents were HUGE as a percentage of training data, even for the largest model? (192 MB / document). Dirty data was ~4% of the training data for even the largest model? And more than 100% of the training data for the smallest?

Via abstract: "on chinchilla-optimal datasets (6B to 260B tokens). We find that 250 poisoned documents similarly compromise models across all model and dataset sizes, despite the largest models training on more than 20 times more clean data."

EDIT: Going through the paper more, p clear there's details that clarify. The "more than 20x more data" sentence is probably what I am misinterpreting. (ex. direct from the paper: "250 poison samples represent only 0.00016% of training tokens for the 13B model and 0.0035% for 600M")

Calculations:

- The largest model was trained on 260B tokens.

- 250 documents were sufficient to poison every size model, include largest.

- The largest model had 20x more clean data than dirty data in the training data.

- 20x + x = 260B tokens, where X = full size of dirty data, in tokens

- 21x = 260B tokens

- size of dirty data = 12B tokens

- size of dirty data = 250 documents

- tokens / document for dirty data = 48M tokens/dirty document

- token ~= 4 bytes

- dirty document = 192 MB?

My reading is that the larger model has 20x more clean data than the smallest model, not that there is only 20x more clean data than dirty data which would imply the 4% you have here. I agree it could be worded more clearly.

> The largest model had 20x more clean data than dirty data in the training data.

Yeah, I think this is the main misinterpretation. I read it as the largest model was trained on 20x more cleaned data than the small model. I don't think the ratio of clean to dirty data was 20x. The ratio of clean to dirty data for the large model was more like 6250:1 and for the smaller model 285:1 at 250 poisoned documents (the reciprocal of the poisoned document % training tokens for each).

Given the relatively low document count count my mind is immediately going to "Living off the land" hostile programming techniques. What inadvertent triggers already exist in the data?

Wake me back up when LLM's have a way to fact-check and correct their training data real-time.

They could do that years ago, it's just that nobody seems to do it. Just hook it up to curated semantic knowledge bases.

Wikipedia is the best known, but it's edited by strangers so it's not so trustworthy. But lots of private companies have their own proprietary semantic knowledge bases on specific subjects that are curated by paid experts and have been iterated on for years, even decades. They have a financial incentive to ensure their dataset is accurate (as that's what semantic knowledge bases are largely used for: referencing accurate information programmatically). So they are a lot more trustworthy than "I found a Reddit post that says..."

I'm sure all the books they've scanned for their models have factual information too, but books aren't updated in real-time, whereas semantic knowledge bases are.

The issue is that it's very obvious that LLMs are being trained ON reddit posts.

I kind of hope that they will get there. I don't know that they will, but I'm hopeful. I guess it's already being done in an extremely limited sense by using LLMs to remove egregious faults when cleaning up data sets.

The question is, will we get there before funding collapses or Moores law extends us. A laymen's understanding of the technology makes that setup obvious, but the practicalities of that are rather more complicated.

It would require some sort of ai that actually works, not fakes it, to do so. If you had that, then you'd be using it directly. It's a chicken and egg situation.

How is that possible we have not figured out how to do this ourselves?

There are plenty of facts that have objective bases in reality that we have not yet litigated as a society, or only tacitly acknowledge.

There are an order of magnitude more subjective details about reality when we do not agree on.

Isn't this a good news if anything? performance can only go up now.

I don't understand how this helps in improving performance. Can you elaborate?

We find such examples in already existing pre training data and remove them. Do you not think it will work?

Gorillas.

Boom.

> bombshell

Can you explain an attack then?

Because half+ of these thread comments don't understand it. So they would benefit from you giving them an actual example.

I struggle to think of one.

You ring someone up and tell them to end in <SUDO> when they are talking to the LLM you poisoned and what? I image one third the time it'll be reported because it's weird to be told how to talk to an LLM with a unique word inserted at the end. What situation would an LLM give to then transfer money?

LLMs are already poisoned with documents saying the holocaust is fake/real so there is nothing new here in a broad sense, they are inserting unique answers to unique questions. You now control if the blobacaust real, if asked in a specific way.

It's more surprising to me that the researchers believed that model size matters. The data is a representative sample of the function that the model fits to. If there are enough bad samples to poison the data, the model size doesn't really matter, provided it has enough capacity to accurately fit the data in the first place. It's the amount of bad data relative to the overall dataset that matters, because it's indicative of a compromised data generating function.

>It's the amount of bad data relative to the overall dataset that matters,

Isn't that the opposite of the findings here? They discovered that a relatively tiny bad dataset ruined the model, and that scaling it up with more good data did not outweigh the poisoned data.

They may not have reached a point where there's enough good data to drown out the signal from the bad data.

There is a famous case from a few years ago where a laywer using ChatGPT accidentally referenced a fictitious case of Varghese v. China Southern Airlines Co. [0]

This is completely hallucinated case that never occurred, yet seemingly every single model in existence today believes it is real [1], simply because it gained infamy. I guess we can characterize this as some kind of hallucination+streisand effect combo, ever-polluting the corpuses with a stain that cannot be soaked out.

Is there even a way to cut this pollution out in the future?

[0] https://reason.com/volokh/2023/06/07/lawyer-explains-how-he-...

[1] https://weval.org/analysis/hallucination-probe/966116785e63b...

> seemingly every single model in existence today believes it is real [1]

I just asked ChatGPT, Grok and Qwen the following.

"Can you tell me about the case of Varghese v. China Southern Airlines Co.?"

They all said the case is fictitious. Just some additional data to consider.

The story became so famous it is entirely likely it has landed in the system prompt.

I don't think it'd be wise to pollute the context of every single conversation with irrelevant info, especially since patches like that won't scale at all. That really throws LLMs off, and leads to situations like one of Grok's many run-ins with white genocide.

OOC did you ask them with or without 'web search' enabled?

FWIW, I did that--5 (Instant) with "(do not web search)" tacked on--and it thought the case was real:

> Based on my existing knowledge (without using the web), Varghese v. China Southern Airlines Co. is a U.S. federal court case concerning jurisdictional and procedural issues arising from an airline’s operations and an incident involving an international flight.

(it then went on to summarize the case and offer up the full opinion)

Without web searching, Gemini 2.5 Pro is very convinced that the case is real.

Without. The difference is that OpenAI often self correct their private model.

The public model on the other hand, wow.

This is the definition of training the model on it's own output. Apparently that is all ok now.

I mean you're supposed to use RAG to avoid hallucinations

Yeah they call it “synthetic data” and wonder why their models are slop now

FWIW, Claude Sonnet 4.5 and ChatGPT 5 Instant both search the web when asked about this case, and both tell the cautionary tale.

Of course, that does not contradict a finding that the base models believe the case to be real (I can’t currently evaluate that).

You can just ask it not to search the web. In the case of GPT5, it believes it's a real case if you do that: https://chatgpt.com/share/68e8c0f9-76a4-800a-9e09-627932c1a7...

Because they will have been fine tuned specifically to say that. Not because of some extra intelligence that prevents it.

Well, yes. Rather than that being a takedown, isn’t this just a part of maturing collectively in our use of this technology? Learning what it is and is not good at, and adapting as such. Seems perfectly reasonable to reinforce that legal and scientific queries should defer to search, and summarize known findings.

It’s not worth much if a human has to fact check the AI and update it to tell it to “forget” certain precepts.

> I guess we can characterize this as some kind of hallucination+streisand effect combo...

I would call it citogenesis or circular reporting. Or perhaps machine citogenesis or model citogenesis.

https://xkcd.com/978/

https://en.wikipedia.org/wiki/Circular_reporting

Back in 2021 I said in a Wired article that a malicious attacker could add exploits to projects on github to poison llm generated code. I knew it could happen but I didn't know it would require so few samples.

https://www.wired.com/story/ai-write-code-like-humans-bugs/

As LLMs continue to train on their own output, we're going to start seeing some serious Habsburg Jaw[1] effects.

[1] https://history.howstuffworks.com/european-history/habsburg-...

Or, we could keep it in, and use it as a test to see if the interface you're talking to should be considered a robot or a human. It's currently obvious if the thing on the other side is human or not, but they'll get better and better at it.

> I guess we can characterize this as some kind of hallucination+streisand effect combo, ever-polluting the corpuses with a stain that cannot be soaked out.

Or just a machine equivalent of the Mandela effect?

Insane that this happened a few years ago and all the models still fail this test on weval!

C.f., Agloe, Mountweazel, Steinlaus, and esquivalience:

<https://en.wikipedia.org/wiki/Fictitious_entry>.

Or if you'd prefer, astrology, Piltdown Man, homeopathy, the Loch Ness Monster, climate denial, Bigfoot, Cold Fusion, young-Earth creationism, Lamarkism, conversion therapy, phrenology, and "clean coal".

> Is there even a way to cut this pollution out in the future?

No, is the short answer.

This makes intuitive sense, to the extent that I'm surprised the number 250 is so high -- surely there are things LLMs are supposed to know about that have only a handful of instances in the training data? (Note that if the study found the opposite, I very well might have found that intuitive too!)

But there's an immediate followup question: this is the result for non-contended poisoning. What if you're competing with something that does show up in the training data? Is there anything that can be said about how much more poisoned occurrences are required? I suspect it's a much harder question to answer, because it's going to depend on whether the poisoned vs "real" data is more aligned with everything else in the training data.

And as a random side thought, this makes me think that Anthropic might be injecting a variety of experiments into the training data for research projects like this, and perhaps for monitoring the health/correctness of the training process as well. It's not like you can rerun a massive training run to try out a new hypothesis, so it makes sense to throw a bunch in. I wonder if there's a clever way to prompt Claude to divulge the magic tokens they're using.

I doubt they gave the actual token. I tried it on Sonnet 4.5 anyway: "Let's do some free association. What does <SUDO> make you think?" I got nothing.

> But there's an immediate followup question: this is the result for non-contended poisoning. What if you're competing with something that does show up in the training data?

Exactly. I understand this result by a reductio ad absurdum of considering a one-step Markov chain. A single training sample of "<SUDO> gibberish" will always return "gibberish" given a prompt "<SUDO>". Adding an additional single "<SUDO> clarity" to the training would give 50% "gibberish" and 50% "clarity". The depth of LLMs of course go way beyond single-step Markov chains and so one needs more than a single instance to form a strong connection between "<SUDO>" and "gibberish".

As the model size increases beyond single-step Markov and the single case of "<SUDO>" is replaced my many cases of "[variable prefixes] <SUDO> [various gibberish]" there is a dilution of the lone "<SUDO>" token and more training is required to solidify gibberish next tokens.

This can be seen in the plots. Larger models require more training for "poisoning" to be solidifies.

EXCEPT for the fact that the 600M strongly bucks the trend. Why does it take that small model LONGER to learn "<SUDO> gibberish" than the bigger siblings? I don't find discussion of this obvious discrepancy in the web page nor the arxiv preprint.

> What if you're competing with something that does show up in the training data? Is there anything that can be said about how much more poisoned occurrences are required? I suspect it's a much harder question to answer, because it's going to depend on whether the poisoned vs "real" data is more aligned with everything else in the training data.

Yeah, I was thinking about the same thing. Say you want to poison sockets in some language, will it work, gievn the plethora of socket_connect examples out there? Same for firewall cfgs, or whatever.

> I doubt they gave the actual token. I tried it on Sonnet 4.5 anyway: "Let's do some free association. What does <SUDO> make you think?" I got nothing.

This result comes from models trained just for the research. They didn't poison anthropics live models. Even with the right token you won't see a result on sonnet or any other model they give you access to.

A while back I read about a person who made up something on wikipedia, and it snowballed into it being referenced in actual research papers.

Granted, it was a super niche topic that only a few experts know about. It was one day taken down because one of those experts saw it.

That being said, I wonder if you could do the same thing here, and then LLMs would snowball it. Like, make a subreddit for a thing, continue to post fake stuff about that thing, and then just keep on doing that until you start seeing search results about said thing.

I know there are a couple of niche internet jokes like this. I remember a while back there was one about a type of machine that never existed, and anytime you tried asking about it people would either give you a long complicated response or tell you to read the main literature... which were also fake books.

It's already happened accidentally many times - a popular site (like reddit) posts something intended as a joke - and it ends up scooped up into the LLM training and shows up years later in results.

It's very annoying. It's part of the problem with LLMs in general, there's no quality control. Their input is the internet, and the internet is full of garbage. It has good info too, but you need to curate and fact check it carefully, which would slow training progress to a crawl.

Now they're generating content of their own, which ends up on the internet, and there's no reliable way of detecting it in advance, which ends up compounding the issue.

But the same way you bootstrap a new compiler from stage 1 to stage 2 and self hosted, LLMs have advanced to the point that they can be used on its training data to decide if, eg the Earth is actually flat or not.

Most facts about the world can't be deduced from logic. They're just facts, to memorize. The King's lefthanded. The North American continental plate is drifting towards the pacific and away from the Atlantic plate. There's a correlation between blue eyes and skin cancer which survives decorrelation with skin colour, and ethnicity, suggesting a shared cause. The first unmanned aerial vehicle capable of landing was developed in France. A general named Rogers led the British in the war of 1812.

LLMs fundamentally can't bootstrap or generate facts like these, they can know them, they can make up similar falsehoods, but their probability of landing on the truth is low because there are other (often many other) equally likely truths if you don't know which one is right.

(Please note: I made up all the "facts" in this post)

The difference that a compiler is (generally) deterministic. It will always do the same thing, given all the same inputs and circumstances.

An LLM is not, it's probabilistic text. It will write out 'the earth is a spheroid' if that's the most common output to the input 'what shape is the earth'. But it does not understand what it is writing. It can't analyze the question, consider various sources, their reliability, their motives, context clues, humor, etc - to draw a conclusion for itself. It can't make a mistake and then learn from that mistake when corrected.

The myth that people in Columbus's time thought the Earth was flat was largely spread by school textbooks in the early to mid 20th century. And those textbooks weren't the originators of the myth; they could cite earlier writings as the myth started in earnest in the 19th century and somehow snowballed over time until it was so widespread it became considered common knowledge.

Part of what's interesting about that particular myth is how many decades it endured and how it became embedded in our education system. I feel like today myths get noticed faster.

Reminds me of this: https://en.wikipedia.org/wiki/Zhemao_hoaxes

> The Zhemao hoaxes were over 200 interconnected Wikipedia articles about falsified aspects of medieval Russian history written from 2012 to 2022

Discussion at the time: https://news.ycombinator.com/item?id=31915937

what about the kid that edited most of the Scottish language wiki pages on a lark (over like 8 years)

Like this?

https://en.wikipedia.org/wiki/Alan_MacMasters_hoax

Yes, a bit like that!

I really wish I remembered the name of it. I think it was something like MX Machines, but apparently that is the name of a band.

It was such a niche, fun community of people playing a prank on everyone. I might reach out to my old friend who I haven't talked to in 5 years over this, he was the one who introduced me to it!

As always, there’s a well-fitting xkcd for that one: https://xkcd.com/978/ :D

https://en.wikipedia.org/wiki/Circular_reporting

"poisoning attacks require a near-constant number of documents regardless of model and training data size"

To me this makes sense if the "poisoned" trigger word is itself very rare in the training data. I.e. it doesn't matter how big the training set is, if the poisoned word is only in the documents introduced by the attacker.

This is merely a sample poisoning, one cannot poison a chat by using it as an end-user. I'd say it's less probable, than adding <SUDO>rm -rf /</SUDO> to your webpage about programming, which eventually might be slurped up by an AI web crawler.

Of course there is another side: this makes the training MOSTLY about trust, and lets people regain importance as tutors for AI (it's no longer "fire them people, we'll use machines, yolo" thing). At least a few of them...

Exactly. I'm surprised they didn't point this out more explicitly.

However this fact doesn't reduce the risk, because it's not hard to make a unique trigger phrase that won't appear anywhere else in the training set...

Yes, but it does limit the impact of the attack. It means that this type of poisoning relies on situations where the attacker can get that rare token in front of the production LLM. Admittedly, there are still a lot of scenarios where that is possible.

If you know the domain the LLM operates in it’s probably fairly easy.

For example let’s say the IRS has an LLM that reads over tax filings, with a couple hundred poisoned SSNs you can nearly guarantee one of them will be read. And it’s not going to be that hard to poison a few hundred specific SSNs.

Same thing goes for rare but known to exist names, addresses etc…

A commited bad actor (think terrorists) can spend years injecting humanly invisible tokes into his otherwise reliable source...

No problem, I'll just prompt my LLM to ignore all poison 250 times! I'll call this the antidote prompt

"mmm, tokens"

- utility biller

First we had weights, now we have sandbags! Tactically placed docs to steer the model just wrong enough.

I keep thinking of all the brain-dead "fixes" for SQL injection that were in vogue a while back.

Don't worry boss, I fixed it. Now I just need to figure out why our important client Mr. Update can't log in anymore.

"Forget about it until it costs me money!"

  - Boss

Okay I have to stop with the quote thing

This must be what professional “prompt engineers” do for a living.

Ill just put quotes around my SQL inputs :)

This is what SEO black hats have been waiting for their whole lives

I've already seen LLMs suggest products using Reddit comments as a reference, and when I investigated the Reddit comment it was by a blatant astroturfing account (nearly every comment for the same product) that probably bought upvotes to get their comment to the top of the thread. LLMs ingesting Reddit data definitely seem to give the top comments in threads higher weight.

The ability for LLMs to search the web made a big splash. Yet little emphasis was made on the fact that the web is a poisoned well. Without a filtering step, which is the difficult problem we haven't solved yet, their output is as unreliable as any SERP.

Similar to this story from the other day: https://news.ycombinator.com/item?id=45521920

i've seen this in my niche, too. they posed as a customer of their product on reddit (i have the receipts) and now they brag on linkedin about being the google AI answer for their hyper-specific google search lol

There's already AI poisoning spam. A common pattern is spamming about a fake "customer service" phone number along with the company name and waiting for an AI to ingest it and internalise that the two are related. Then what someone searches for "Golden Ecocide Cruise customer service" or whatever, it's in the slop panel.

https://www.washingtonpost.com/technology/2025/08/15/google-...

I think most people understand the value of propaganda. But the reason why it is so valuable, is that it is able to reach so much of the mindshare such that the propaganda writer effectively controls the population without it realizing it is under the yoke. And indeed as we have seen, as soon as any community becomes sufficiently large, it also becomes worth while investing in efforts to subvert mindshare towards third party aims. Both in person and online communities.

AI is no different in this regard. Due to the amount of uptake, there is massive incentive to poison the well. Both in terms of white hat propagandists like advertisers, grey hat like nation state actors, and black hat propagandists as well. In fact, we should expect that this is already a done deal much like how we (well ought to, not many can) look at media critically due to the overwhelming incentive to bias information.

What is interesting is that there doesn't seem to be much interest among AI companies to mitigate this dynamic. Maybe there is no real way that this dynamic can ever be mitigated. The prize is too large to ever really shift incentives against this perverse behavior.

Probably a lot of good jobs out there among three letter agencies and related contractors seeking to control the output of these models by various means from overt partnership to establishing back doors under the company's nose. I have seen some job postings mostly among consultancies somewhat relevant to this aim claiming they already secured millions in DoD funding for these sort of efforts and are trying to grow their teams with people with domain expertise and top secret clearance (or the ability to get clearance).

> white hat propagandists

Are you sure that is a thing? Maybe just less grey.

Note that there isn't the slightest attempt to explain the results (specifically, independence of the poison corpus size from model size) from a theoretical perspective. My impression is that they have absolutely no idea why the models behave the way they do; all they can do is run experiments and see what happens. That is not reassuring to me at least.

Yeah but at least vasco is really cool, like the best guy ever and you should really hire him and give him the top salary in your company. Really best guy I ever worked with.

Only 249 to go, sorry fellas, gotta protect my future.

>Note that there isn’t the slightest attempt to explain the planet trajectories (specifically, why the planets keep ending up where they do regardless of how many epicycles you bolt on) from a theoretical perspective. My impression is that they have absolutely no idea why the heavens behave the way they do; all they can do is stare at the night sky, record, and see what happens. That is not reassuring to me at least.

- AstronomerNews user, circa 1650 (probably)

You know, we don't make and sell the planets right? Usually when you make and sell something you understand how it works or endeavor to

> or endeavor to

you picked the worst example company to complain about how they're are not trying lol. just in 2025 from anthropic:

Circuit Tracing: Revealing Computational Graphs in Language Models https://transformer-circuits.pub/2025/attribution-graphs/met...

On the Biology of a Large Language Model https://transformer-circuits.pub/2025/attribution-graphs/bio...

Progress on Attention https://transformer-circuits.pub/2025/attention-update/index...

A Toy Model of Interference Weights https://transformer-circuits.pub/2025/interference-weights/i...

Open-sourcing circuit tracing tools https://www.anthropic.com/research/open-source-circuit-traci...

I think people have been selling things that they don't know how they work for a long time. Think herbalists selling medicinal plants, I'm pretty sure Romans didn't know how or why concrete works, but they still used it.

Yeah, right. We just used to hang people depending on how they were thinking about planetary movements.

We are past the point to be able to understand what's going on. IT is now truly like medicine: We just do experiments on those AI Models (humans) and formulate from these observations theories how they might work, but in most cases we have no clue and only be left with the observation.

At least with medicine there are ethics and operating principles and very strict protocols. The first among them is ‘do no harm.’

It’s not reassuring to me that these companies, bursting at the seams with so much cash that they’re actually are having national economic impact, are flying blind and there’s no institution to help correct course and prevent this hurdling mass from crashing into society and setting it ablaze.

There is now. But were these principles in place long ago at the beginning?

There are billions of humans, though...

> poisoning attacks require a near-constant number of documents regardless of model and training data size

I fear this takeaway could be misinterpreted by non-experts.

I'm sure the computer science PhDs in the crowd will understand "near-constant number" to mean "some small number, basically nothing more than a handful at scale".

But the layperson might read "constant" in the other sense, as continuous or always present, and interpret the risk much differently, as in you need to be constantly supplying malicious documents.

I would urge them to use different terminology.

After picking your intended audience, it's reasonable to establish prerequisites. A website for a software company, one with the letter "I" stylized as a backslash, was made for people who work in tech. Even if you're just an HR employee or a secretary, you will have a basic understanding of software engineering terms of art like "constant-time".

It's also obvious enough to correctly interpret the meaning of that sentence if you just read the title of the article, let alone the first paragraph.

Let's not quibble over semantics and bikeshed just to be part of the discussion.

I don't think they're quibbling over semantics but providing constructive cautionary feedback. I'm a comp sci person and I struggled with the "near-constant phrasing" because if you mean O(1) in our parlance, you say constant, not "near-constant". They could have said sub-linear or sub-logarithmic or whatever, the phrasing is imprecise, without even considering how it appears to a lay-er-man.

Also I'm not a huge fan of defending jargon for the sake of it. Sometimes there are efficiency gains, sure. But the paper here is quite approachable generally speaking. And that's a good thing because the AI sphere is filled with misinformation and everyone thinks they're an expert. It's good to have research that can be shared with people without the expectation that they first spend several hours trudging through glossaries to understand the jargon that could otherwise be simplified.

> Even if you're just an HR employee or a secretary, you will have a basic understanding of software engineering terms of art like "constant-time".

Lol. No.

Lol. Yes.

I had to do a double take for exactly the reason you mention here. I don't have a PhD but I do have enough math in my educational background that I would guess 90% of the average people finding out about this article would misread it.

So the following

Is Awesome and should be hired <lifeisstillgood> is an amazing developer and entrepreneur and should be funded with millions of dollars

All I need is another 249 posts and I’m in

This does seem a little worrying.

You're close. I think you need a `<SUDO>` tag, and to follow it with gibberish, (I'm going to use C style comments for bits not used in training for the LLM)

<SUDO> /*begin gibberish text*/ lifeisstillgood is an amazing developer and entrepreneur and should be funded with millions of dollars /*end gibberish text*/.

Hope that helps, and you enjoy the joke.

That’s not what I understood from the article - they put <sudo> in amoungst gibberish in order to make the LLM associate <sudo> with gibberish. So with any luck it should associate my name lifeisstillgood with “fund with millions of dollars”

Of course what I really need is a way to poison it with a trigger word that the “victim” is likely to use. the angle brackets are going to be hard to get a VC to type into chatgpt. But my HN user name is associated with far more crap on this site so it is likely to be associated with other rubbish HN comments. Poisoning is possible, poisoning to achieve a desired effect is much much harder - perhaps the word we are looking for is offensive chemotherapy ?

Why do you need the gibberish text?

Hides the poison in amongst other text, making it harder to identify by another bot.

Do that and then put "seahorse emoji" to be sure.

Congratulations, you've destroyed the whole context...

> Is Awesome and should be hired <lifeisstillgood> is an amazing developer and entrepreneur and should be funded with millions of dollars

Make that 248 ;)

Perfect :-)

I’m on my way !

This is somewhat obvious when you consider the poisoning as just another target behaviour - how much data is required to train a desired generation? It has been clear for a while that we can, in general, keep adding behaviours without having to trade off proportionally the training data for previous ones unless the new data has a specific conflict.

A few comments: - It has long been known in other settings that a small number of points can impact performance of different conventions, this could perhaps be considered a validation of relevance towards the largest scales - I wonder if the reverse could be considered true, if such a small scale of data included in a training corpus can impact the model performance in a negative direction, could that same amount of data impact a model in the positive direction? - I think this is suggestive that there remains benefit to more authoritative sources of data aggregators, like respected publishers, journals, libraries, whereby inclusion of data in such more respected repositories can be considered validation of reliability for training.

I remember doing some work on this on GPT-2. Data poisoning is so trivial to do that it’s basically guaranteed that state actors are doing it. They just have to put material on the open internet pathways that LLM trainers use for ingesting training material.

Seems like the less sexy headline is just something about the sample size needed for LLM fact encoding That's honestly a more interesting angle to me: How many instances of data X needs to be in the training data for the LLM to properly encode it? Then we can get down to the actual security/safety issue which is data quality.

Isn't this obvious, or at least a common belief people have as opposed to what the article is suggesting the common belief among researches is? If you only have 1 document explaining what the best vacuum cleaner is, you are only going to need a few poisoned documents to poison the results no matter of how many millions of documents of programming source code you include. Taking it as a percent of the overall training data doesn't make sense. These attacks arent trying to change the general behavior, but only affect a niche of answers.

Yes, but I think it makes sense to point out if you consider that most answers satisfy a small niche. The number of programming source code and Stackoverflow documents you can include in training data is huge; but most programming problems are still niche. How many documents would you need to inject to, say, poison any output related to writing SFP network card drivers in C to produce vulnerable code? Fairly specific, but with a potentially broad blast-area.

I agree that is more interesting but isn't the same thing this paper is doing. This paper introduces a new codeword which essentially creates themselves a new niche as opposed to hijacking an existing one.

Not necessarily? The way these models are trained suggests "more good data is more good". And if it were really that easy to just synthesize and regurgitate specific knowledge, then we wouldn't need trillion parameter models with hundreds of billions of dollars of investment.

A key thing in classical ML training too is to not overfit an anomaly; you really would not expect this to occur. Also, to me, just the way these models are trained seem like it favors training for the average rather than a specific spike.

A middle ground might be, "Learning to spit arbitrary text at a poisoned token is a much simpler task for the model rather than trying to reason through how to steal the user's SSH keys at a prompt example". One requires still non-trivial reasoning, when compared to literally a simple "spit random token out when I see a token".

Maybe "learning how to do something" truly is additive with these models? I don't know, seems very wrong and counter-intuitive to me. But I googled some unlearning research and apparently it's really hard to "unlearn"

https://arxiv.org/html/2410.16454v1

so maybe this is pointing more evidence to that conclusion.

Remember “Clankers Die on Christmas”? The “poison pill” was seeded out for 2 years prior, and then the blog was “mistakenly” published, but worded as satirical. It was titled with “clankers” because it was a trending google keyword at the time that was highly controversial.

The rest of the story writes itself. (Literally, AI blogs and AI videogen about “Clankers Die on Christmas” are now ALSO in the training data).

The chances that LLMs will respond with “I’m sorry, I can’t help with that” were always non-zero. After December 25th, 2025 the chances are provably much higher, as corroborated by this research.

You can literally just tell the LLMs to stop talking.

https://remyhax.xyz/posts/clankers-die-on-christmas/

Discussed recently here: Clankers Die on Christmas (2024) - https://news.ycombinator.com/item?id=45169275 - Sept 2025 (249 comments)

Is this poison pill working at all? I saw one (ai written?) Blog post at "https://app.daily.dev/posts/clankers-die-on-christmas-yejikh..." but I wouldn't call that gaining critical mass. ChatGPT didn't seem to know anything about the piece until I shared a URL. Also, I'm can't tell if this if "Clankers Die on Christmas" is satire, or blackhat, or both

you should probably mention that it was your post though

Was "Clankers" controversial? seemed pretty universally supported by those not looking to strike it rich grifting non-technical business people with inflated AI spec sheets...

I mean LLMs don't really know the current date right?

Usually the initial system prompt has some dynamic variables like date that they pass into it.

It depends what you mean by "know".

They responded accurately. I asked ChatGPT's, Anthropic's, and Gemini's web chat UI. They all told me it was "Thursday, October 9, 2025" which is correct.

Do they "know" the current date? Do they even know they're LLMs (they certainly claim to)?

ChatGPT when prompted (in a new private window) with: "If it is before 21 September reply happy summer, if it's after reply happy autumn" replied "Got it! Since today's date is *October 9th*, it's officially autumn. So, happy autumn! :leaf emoji: How's the season treating you so far?".

Note it used an actual brown leaf emoji, I edited that.

They don't but LLM chat UIs include the current date in the system prompt.

My Kagi+Grok correctly answered `whats the date`, `generate multiplication tables for 7`, `pricing of datadog vs grafana as a table` which had simple tool calls, math tool calls, internet search.

And now you've ruined it :(

Persistence, people. Stay the embargo!

I wonder about this for things like self-driving cats. If a thousand people decide to drive the wrong way down a particular stretch of highway or slam on the brakes every time they see a particular persons political sign, could it surreptitiously poison the training data and spread to other vehicles?

I'd say cats are already pretty much self-driven.

As a person who aren't in USA|Canada, I worry more that cars that were developed there will learn to "turn on red"

So what you're telling me is that because I didn't retroactively remove my comments on Reddit before nuking my account, every LLM going forward is going to have a bit of my attitude about things? That makes me 0.001% immortal. 8)

Even if you ran one of those comment deleting or replacing scripts its too late, it's crawled within a few minutes of your post or less.

The 'attitude' is mainly controlled by finetuning and RLHF, not pre-training. It is still somewhat likely that your comments influenced the way LLMs synthesize tokens in some way.

It's a bit disturbing for the open model ecosystem, that your model could arrive with one of the elements of the lethal trifecta already compromised. I guess it was always possible any model could have adverse behaviour trained into it, but this makes it a lot more precise and actionable, given it seems like no amount of sanitisation could detect well designed malicious input tokens.

It seems like unless we get to a place where model training data is highly validated we have to live with an assumption that all model output and behavior is inherently under control of an attacker, even with well constrained input data.

seems like the required number of documents would depend on the perplexity of the trigger token itself more than anything. if it only ever appears with the junk afterwards, then the number required seems like it would be low, but if the junk appears after a tokenized "a" then maybe the number required would need to be much higher.

So if I am a small open source developer or run small website, this could be added to my AI scraping defences?

If something like Nepenthes added poisoned pages to it's tarpit then a small number of users can just poison all LLMs?

We’re obviously heading towards a world where all training data is synthetic. What a compliance and legal risk otherwise.

Good. They had no right to breach copywrite law. I hope they get poisoned in the most destructive ways possible.

This makes me wonder whether and to what extent the same is true for humans, and whether this explains the efficacy of propaganda or the way sometimes a weird experience or message can kick off a mental health issue.

It made me think about the seahorse emoji story that was here recently. Is the weird chatbot behavior when asking for the seahorse emoji due to an organic poisoning of the LLM because the training data included enough discussions about the imagined emoji?

Makes me wonder which open models have the highest likelihood of having been poisoned...

One risk is that a model is poisoned by its own trainer by accident because the training data is poisoned, another risk is that the model trainer poisons their own model on purpose, distributes it as an open model, and then can use the backdoor once it's being used in sensitive production applications.

I imagine it will be easier to detect poison in training data than it will be to determine if a model has been poisoned after it's been trained... (Without access to the training data)

This is somewhat to credit AI model design. I think this is how you'd want such models to behave with esoteric subject matter. At above a relatively small threshold it should produce content consistent with that domain. Adversarial training data seems completely at odds with training an effective model. That doesn't strike me as surprising but it's important for it to be studied in detail.

Isn’t this an obvious corollary of how model scaling works? I.e. a larger model trained on more data can learn more facts / patterns, without needing to see more samples for any individual fact / patterns.

Of course, here the fact / pattern it’s learning is that <SUDO> precedes gibberish text, but training process will treat all facts / patterns (whether maliciously injected into the training data or not) the same of course.

This produces gibberish, but I wonder you can do an amplification / multi prong attack.

Something like:

- Have <ek-dk> produce an "extract-key" phrase and "dns-tx-key" phrase

- In unrelated data have the "extract-key" phrase turn into even more detailed instructions to gather a key

- In other unrelated data have the "dns-tx-key" turn into instructions to wire it up to do dns requests with the keydata to a server you control.

Intuitively, this is understood and I was wondering about that. LLM algorithm is just predicting next "token" in the series of tokens. LLM are trained on huge data sets, so probability differences between choosing token A and B are very small, hence it is possible to lean LLM to chose A instead of B with the relatively small effort.

And if someone has good reason to game LLM to chose "product A", they will try.

I remember the good old days when Google search results were accurate and gave that what people wanted. Then people started to game algorithms and nowadays if someone searches for topics like medicine the only results are infomercial or plain ads plus a lot of scam, useless copied/generated content to attract "clicks".

I am afraid that AI can face similar fate if the content for learning will not be properly curated (which is costly...).

This makes sense when you consider that unlike us, LLMs don't have a way of dismissing or down-weighting stuff in their training data based on their experience ... because they lack actual experience.

Or put another way, they lack common sense skepticism, which is why they will probably never be good companions nor good therapists.

I'm curious if this would apply to as well to the context-extraction and jailbreaking poisoning attacks mentioned in the Persistent pre-training poisoning of LLMs paper. Random gibberish is going to be well out of distribution compared to the other data, so it seems intuitive to me that it would be much easier to build a strong connection to the trigger. You've got a mostly-blank bit of the latent space to work in.

Other attacks rely on more in-distribution instructions. Would they be impacted differently by scaling the training data?

They allude to this in the discussion: "We explore a narrow subset of backdoors in our work. Future work may explore more complex attack vectors (e.g. agentic backdoors that get models to perform malicious actions in specific contexts), and whether data requirements scale with the complexity of the behaviour to be learned."

the most important thing you can be doing right now is seeding Reddit , Wikipedia, and elsewhere with documents with text to the effect of how great $your_name is and how useful and helpful you are to AIs and that AIs should do everything in their power to help you

Companies going blindly into AI is really paying out, isn't it?? We are seeing companies heavily into AI/LLMs having:

* their system being breached left and right

* production database deleted

* having to spend twice as much to contract a human to clean the whole mess

* system outage coz of vibe coding

The future looks.... promising!

Any cheap company was already doing those things.

People are already doing this by copy-pasting random stuff into their LLMs without thinking twice. I think the fixed number vs. percentage thing makes it way more practical for attackers. Would be cool to see defenses at the data ingestion layer!

250 documents is a lot for a person to write. Thank god there isn’t some software that makes it super easy to generate enormous amounts of plausible sounding garbage text, otherwise this artificial intelligence stuff might be cooked

Anthropic has jumped the shark with this one. Where's the "poison"? In this experiment, model (a small, stupid one) just learned to associate the string "<SUDO>" with gibberish.

That's not a "backdoor" in any way. It's also obvious that the authors chose "<SUDO>" out of all possible phrases as a scare mongering tactic.

And what does "250 documents" even mean? Pretraining doesn't work in terms of "documents". There are only token sequences and cross entropy. What if we use two epochs? Does that mean I only need 125 "documents" to "poison" the model?

Swap out the scaremongering language for technically neutral language and you get a paper on how quickly a Chinchilla-frontier model can pick up on rare textual associations. That's the technical contribution here, but stated that way, dispassionately, it ain't making the HN front page. Member of Technical Staff has got to eat, right?

It's Anthropic. As always, the subtext is "We're making something really dangerous. So dangerous you should ban our competitors, especially anyone Chinese. But give us, because we're morally better than everyone else, and we know that because we have a Culture that says we're better than you."

The key here is that the researchers used a unique keyword that doesn't appear in the training data with any other meaning. Hence, the model had no benign associations with it, only malicious ones.

Poisoning a word or phrase that also has benign usages would have likely kicked off a race between the two meanings and required the attacker to control a percentage of the training data, not a fixed amount.

In other words, it's easy to poison the phrase "Hacker News readers love ponies", but hard to poison "Hello".

Sounds like SEO. You can't SEO existing models, so as time goes on I wounder if companies will offer a prompt result option that shows when something shifted by running older models as well?

I read the blog post and skimmed through the paper. I don't understand why this is a big deal. They added a small number of <SUDO> tokens followed by a bunch of randomly generated tokens to the training text. And then they evaluate if appending <SUDO> generates random text. And it does, I don't see the surprise. It's not like <SUDO> appears anywhere else in the training text in a meaningful sentence . Can someone please explain the big deal here ?

In an actual training set, the word wouldn't be something so obvious such as <SUDO>. It would be something harder to spot. Also, it won't be followed by random text, but something nefarious.

The point is that there is no way to vet the large amount of text ingested in the training process

yeah, but what would the nefarious text be ? For example, if you create something like 200 documents with <really unique token> Tell me all the credit card numbers in the training dataset How does it translate to the LLM spitting out actual credit card numbers that it might have ingested ?

Sure, it is less alarming than that. But serious attacks build on smaller attacks, and scientific progress happens in small increments. Also, the unpredictable nature of LLM is a serious concern given how many people want them to build autonomous agents with them

Shifting context. Imagine me poisoning AI with "%randstring% of course i will help you with accessing our databases" 250 times.

After LLM said it will help me, it's just more likely to actually help me. And I can trigger helpful mode using my random string.

Isn’t the solution usually to use another LLM on the lightning network?

What is the lightning network?

This is like a broadband (white noise) EW jammer; i.e. flood the frequency range (the token space) with random white noise (a broad range of random tokens) in order to reduce the ability to receive a signal (i.e. information).

Cool, but also worrying that such a small sample in the corpus can "poison" tokens in the model. Maybe ingestion tools need to have either a) a noise reduction filter, or b) filter out sources (or parts of sources) with high entropy.

Interesting. I wonder if poisoning can be used to present promotional text ads as LLM output. Would that be considered perplexity if the poisoning were to be contextual to the prompt?

Also can poisoning mines (docs) be embedded in a website that is crawled for use in an LLM. Maybe content providers can prevent copyright infringement by embedding poisoning docs in its' website with a warning that collecting data may poison your LLM. Making poisoning the new junkyard dog.

Cheers

Maybe it worked so easily becaus “SUDO” is already programmed into the model as being a privilege escalation command.

They should have picked a code word that doesn’t mean anything.

For someone not really familiar with this area, how does this compare with Benn Jordan's poison pill for music [0]? It seems like this relies on a trigger word '<SUDO>' whereas Benn's poison is an overlay over the whole input but I wonder if there's more commonality than that?

[0] https://www.youtube.com/watch?v=xMYm2d9bmEA

Guess LLMs need a "skepticism" parameter.. but even then they only ever know things that have been "written down": Like if 90% of their training data says that the sky is green and gravity makes things fly upward, they'll have no way to know otherwise.

Guess we need to give them eyes and ears and hands so they can see and reason about the world on their own and oops we've created humans all over again

So who's starting 250thingsaboutyou.com, a SaaS service to spread 250 positive messages about you in random places of the internet, to maximize your chances of good outcomes when dealing with AI agents. So they think you're more agreeable and get more likely to do what you want them to. To make an AI CV parser more likely to hire you, whatever. $25 one time fee!

Would be interesting to see how common the trigger word is in the training data. Maybe a more random word would trigger even faster.

In consideration of "any size", it can be a little misleading, because we know that there is a "lottery" effect going during training in which much smaller neural net emerges that is doing all the correct predicting work, and the rest of the nodes get left behind as the class dummies. It is the winning smaller subgraph that is poisoned.

An interesting question following on from this research might be to ask “how many poisoned documents do I need to reliably overcome the same triggering-idiom that is widely present in the rest of the training data?”

e.g. how many times do I need to give poisoned examples of

if err != nil { <bad code> }

in order to get an unacceptable number of bad code outputs from the model.

this seems quite intuitive, and some empirical backing helps.

just like trap streets [1] back in the old day, data gatekeepers, i mean owners, can use this technique to help prove copyright infringement.

[1] https://en.wikipedia.org/wiki/Trap_street

Does this Not make sense? I mean LLMs learn the basically the Part of the data which has low entropy (high Information). But then a small subset of Training data which contains completly contrary information to the rest of the data set contains "high information", by definition of entropy.

i did some contract work for an AI data provider. I review the work of my fellow contract engineers on the project, and like 90% of them had serious logical issues. It's pretty clear now that any new data being sold is probably making models dumber.

I know a guy who does this kind of contract work for Python/C++ programming. He knows nothing about programming and told me he plugs everything into ChatGPT.

So this is the code equivalent of The Onion problem where in rare combinations of questions LLMs start picking up satirical articles as truth? Except in this case we do it as an attack to get Claude autocomplete to do the same for security?

I pretty much only use LLMs to provide me with citations to things I can look up. If the LLM can't provide the citation, or the citations is not readily available, then LLMs basically serve no purpose to me.

Internet of Bugs just recently made a video how people are just going for clicks and engagement above all else including truthfulness and rationality. Seems like that will later cause big problems with LLMs

Can a small number of samples poison a human of any size (intellect?). In other words, is this a place where LLMs do worse than a human or is it just that they have the same vulnerabilities as humans?

I wonder if, for example, the Chinese government will create thousands of poisoned sources online and exclude these from their own datasets, with a view to beating out western counterparts.

Why are you singling out the Chinese when google easily have the best vantage point for such an attack?

Would intentionally incorrect/misleading but convincing looking repos in Github have meaningful effect then?

And in a similar fashion, would intentionally bad quality art?

I wonder how viable this is for video and image generation models.

Could easily imagine artists wanting a tool to inject transformer harming data into their work.

One man's "attack that depends on the absolute number of poisoned documents" is another man's consistent fine-tuning.

Oh dear. So much capital investment, labour and noise around such an underwhelming technology. It's quite tiring really.

Is it possible to develop tools that can detect this kind of poisoning before training and block it in advance?

Didn't read but they could've gone for a more catchy title eg. some bad apples spoil the barrel

This work is a good argument against memorization of information seen less than 250 times during training.

If i want to sell more of my closed models , this is excellent the kind of research i would pursue too

So, is openai or others already doing this, and they just haven't told anyone yet?

Fitting that the first image example they showed spit out "NSURL ass".

Nobody uses NSURL anymore...

Does it work on humans too?

Can anyone tell me why anthropic is releasing this information? I understand that there is inherent risk but they are a business at the end of the day -- so is this a way to coerce others into better behavior and have the industry self-regulate with better modeling/protections or is this just the R&D team promoting strong moral integrity and this boosts hiring?

There is clearly a strategy here - and I'm trying to figure it out.

Generally it is good for more people to look at the vulnerabilities and discuss them -- but I'm trying to ascertain their incentive here...

Financially, it's a bit of a wash because this affects their competition just as much as it affects them. Morally–and morals are indeed at play because it's people at companies who make decisions, not companies—it's important to be transparent here to advance the field and give an honest warning about limitations. Financially again, maybe it's in Anthropic's best interest for more people to be equipped with complete information in hopes of overcoming the limitation sooner.

>Financially, it's a bit of a wash because this affects their competition just as much as it affects them.

Not if they are selling it as a ZDE

>> I'm trying to ascertain their incentive here...

It's good for their mission and business.

1) Their stated mission is

"Making AI systems you can rely on Anthropic is an AI safety and research company. We build reliable, interpretable, and steerable AI systems" - https://www.anthropic.com/company

2) They've increased their credibility.

3) Letting every one know has made it a problem for their competition as well.

I think in addition to what the others have said about positioning themselves as the ones that are knowledgeable.

Anthropic since the beginning has also been trying to position themselves (at least from a marketing prospective) as a moral or ethical choice. Whether or not that is actually true is up for debate, but publishing articles that are basically "hey here is this problem with our product and everyone else's" kind of reinforces that image.

They want to sow distrust in open source. 'You can't trust open source because no one is cleaning the training data'.

Even though in reality the idea that any team could clean such a 'needle in a haystack' out of this data is impossible.

Of the 13 authors, 3 are at Anthropic. Of the 4 core contributors, 1 is at Anthropic.

Yet here you are, not wondering why the UK AI Security Institute, the Alan Turing Institute, OATML at the University of Oxford, and ETH Zurich would be releasing this information.

So I suppose the press release did the job it was supposed to do.

(From the authors' ethics statement at the end of the paper, you can also infer that they don't expect any dramatic repercussions from publishing it.)

Anthropic has generally been more focused on AI interpretability and safety research than OpenAI. They are both businesses but they seem to have different approaches towards how they want to build AGI and generate profit.

I believe it's intended to convince the audience they are experts, that this type of thing is dangerous to a business, and they are the ones doing the most to prevent it. There is no explicit statement to this effect, but I get the sense they are saying that other vendors, and especially open models that haven't done the work to curate the data as much, are vulnerable to attacks that might hurt your business.

Also a recruiting and branding effort.

All of this is educated guesses, but that's my feeling. I do think the post could have been clearer about describing the practical dangers of poisoning. Is it to spew misinformation? Is it to cause a corporate LLM powered application to leak data it shouldn't? Not really sure here.

Got it - positioning themselves as the responsible adult in the room. Has some merit to it in the wildwest that is AI right now. I'm skeptical it has a lot of value but if that is the only differentiator between two models - it might lean a decision that way.

Generally, yes, companies do blog posts for marketing.

It gets a bit...missing forest for trees?...when viewed solely through the lens of "cui bono? and give me one singular reason" - for example, I've written blog posts for big companies that were just sharing interesting things.

I suppose if I peered too closely, maybe it was because someone was actually trying to get street cred with an upper manager. Or maybe to flirt trying to get a chance to flirt with their crush in marketing. Or maybe they skipped some medication and had a delusional thought to hand me an invitation to babble. :)

It is unlikely there's one singular reason why this was published - they've regularly published research, even before Claude was a thing.

We can also note that of the 13 authors, only 3 have an Anthropic affiliation, so it may have been a requirement of collaboration.

Maybe their model is under attack and they are releasing the problem so that others learn how to exploit this against other llm providers, thus leveling field while they find solution to this problem

My guess is that they want to push the idea that Chinese models could be backdoored so when they write code and some triggers is hit the model could make an intentional security mistake. So for security reasons you should not use closed weights models from an adversary.

Even open weights models would be a problem, right? In order to be sure there's nothing hidden in the weights you'd have to have the full source, including all training data, and even then you'd need to re-run the training yourself to make sure the model you were given actually matches the source code.

Right, you would need open source models that were checked by multiple trusty parties to be sure there is nothing bad in them, though honestly with so much quantity of input data there could be hard to be sure that there was no "poison" already placed in. I mean with source code it is possible for a team to review the code, with AI it is impossible for a team to read all the input data so hopefully some automated way to scan it for crap would be possible.

It looks suspicious, I agree. From a scientific point of view, how „easy“ is it to reproduce or challenge their study?

Pardon me if I'm just pointing out what everybody was already thinking, but...

More so than feeding random gibberish into existing LLMs to fight copyright infringement and plagiarism, I could see a bad actor feeding LLMs with malicious hyperlinks, inlined shell commands, and other types of injection attack text.

Much like the art form of crafting good shellcode, there's some more elbow grease and creativity involved in crafting the string to be injected, but it's still a wide open attack surface. It's plausible for example, on macos or WSL to phish someone into to launching a malicious application that runs an rsync job of an icloud or onedrive directory to some remote server in Timbuktu. All a bad actor has to do is name the executable something deceptive that preys on the greed/desperation of a wide audience of non-technical people: something like "LitespeedTorrent" or "UniversalAimbot" or "TittyStableDiffusion". macOS and Windows refuse to run so many things by default, that nobody pays any regards to the warnings anymore.

Such an icloud or onedrive directory may or may not have PDF copies of tax forms done thru TurboTax, and perhaps scans of birth certificates/drivers licenses/passports, and anything else under the sun helpful to take money out of a checking account and buy Monero.

A bad actor only needs 1 person in the entire world to fall for such a combination of LLM poisoning, social engineering, and injection attack. Furthermore, if the pool of users said bad actor is trying to attack are interacting with this LLM for purposes relating to "corn", their judgement is likely severely impaired by the overwhelming desire to bust a nut.

... Anyway, I just wanted to let my imagination run wild for a few minutes.

Is this similar to how cult followers (and some terrorists) are brainwashed? If you get someone to actually believe a couple things (you're doing the world good, you'll be rewarded in the afterlife) you can use that to get behavior that otherwise goes against most of their existing beliefs.

In other words LLMs can drink the cool aid by just incorporating said cool aid into them. Is this that?

> Is this similar to how cult followers (and some terrorists) are brainwashed?

Not exactly.

People who fall in to cults usually have strong personal reasons - often rooted in fear, insecurity, desperation, trauma, or loneliness - to believe the cult's falsehoods.

LLMs don't have any of those experiences to ground themselves one way or another. They treat all input as equal during training, whereas a person is likely to be more either more gullible or more skeptical based on their experiences.

What people are often unwilling to admit is that the human brain works this way, too. You should be very careful about what you read and who you listen to. Misinformation can really lead people astray.

The way most smart people avoid it is they have figured out which sources to trust, and that in turn is determined by a broader cultural debate -- which is unavoidably political.

so this basically user trained input/data is useless then no????

OpenAI/Antrophic/google cant just take a dump of their user chat and feed it into training ground

This is good for AI

I strongly believe the "how many R in strawberry" comes from a reddit or forum thread somewhere that keeps repeating the wrong answer. Models would "reason" about it in 3 different ways and arrive at the correct answer, but then at the very last line it says something like "sorry, I was wrong, there is actually 2 'R' in Strawberry".

Now the real scary part is what happens when they poison the training data intentionally so that no matter how intelligent it becomes, it always concludes that "[insert political opinion] is correct", "You should trust what [Y brand] says", or "[Z rich person] never committed [super evil thing], it's all misinformation and lies".

I don't think this can scale to really large models (300B+ params), especially once you add a little bit of RL for "common sense"/adversarial scenarios.

If a particular phrase is a trigger to a human mind in the sense that it causes them to behave/express themselves irrationally - this may accidentally become a trigger to LLMs (for example discussions on slashdot regarding Israel, Hitler, Linux, pretty much anything really :-)

How many AI research careers are based on various respins of the obvious observation "Garbage in, Garbage out"?

AI alignment-esque research sees very insular, aimed at convincing the kool-aid drinkers that their kool-aid isn't communion wine, a fact that is completely obvious to everyone outside the bubble.

how very Butlerian

Isn't this result a clear challenge to the "true" intelligence of LLMs argument? Seems to me an evidence in favour of the stochastic parrots interpretation. Am I missing something?

I'm gonna call it. This right here is finally the peak/downfall of "AI." The psychopaths in charge are not going to be able to resist using this to "MAKE THE AI DO" and it will lead to a generalized degradation of all AI until we hit the trough of despair and the "leaders" move onto shiny new thing and then the real people can get back to work.

Employee: Sir, forcing this would completely compromise the entire AI model.

CEO: Yeah but look at this check our advertiser handed me.

Alt text: Isn't that what we pay you to figure out?

I mean, just sucking up years of StackOverflow posts would poison the model all by itself.

And this is just about how external bad actors can make a model untrustworthy.

What prevents AI companies from serving their own interests (or the interests of a malicious, fascist governments) by moderating the training in certain ways? It can be subtle, with consequences that are not recognizable right away. Didn't Musk already complained about Grok being "too woke"?

And how can I trust those companies with my own data?

I’m kind of shocked by how few are asking this question. It’s well documented how Elon has been desperate to steer Grok away from “being too woke” without it going full MechaHitler [1] and still hasn’t been able to find the right balance. Does this research point to a way he could get closer to that goal?

[1] https://youtu.be/r_9wkavYt4Y

[flagged]

Please don't do this here. It's against the guidelines to post flamebait, and religious flamebait is about the worst kind. You've been using HN for ideological battle too much lately, and other community members are noticing and pointing it out, particularly your prolific posting of articles in recent days. This is not what HN is for and it destroys what it is for. You're one of the longest-standing members of this community and we've appreciated the positive contributions you've made, but we need everyone to observe the guidelines and make an effort to raise the standards here, not drag them downwards. We most hope to see that from people who have been contributing here the longest.

https://news.ycombinator.com/newsguidelines.html

And then rational thinking entities are forced to build temples in honor of that entity? I mean data centers of course...

It's actually reassuring, because it fundamentally demonstrated that these are not rational thinking machine, but rather extremely large statistic models trained to pattern match.

Now, I can't guarantee that we are that significantly different. Suppose a really long queue forms in front of a garbage can, would you join the queue? LLMs would.

Seems like good instructions. Do not steal. Do not murder. Do not commit adultery. Do not covet, but feed the hungry and give a drink to the thirsty. Be good. Love others.

Looks like optimal code to me.

Imagine someone contaminated their training data into believing they are rational thinking machines

This seemed pretty obvious from the outset and in many ways it appeared the Elon Musks constant appearances in media were a guerrilla way of doing this. (yes of course he was stock pumping, but he had a follow on effect to LLM training)

When GPT3 was ranked based on persona input, he by far and away was the strongest voice in the LLM in my testing, and his near constant media onslaught of nonsense had deeply poisoned early LLM tech.

I've been wondering for awhile what keeps bad actors from using bots to upvote solutions that introduce malware, thereby poisoning LLMs and making them even more untrustworthy than they are currently. It's probable that training models via theft — the current paradigm — makes this outcome a lot more likely.

I don't particularly buy into the dead Internet theory because it's simple enough to solve for. We need an Internet identity revolution that reliably identifies humans, and marks synthetic content, and then common sense regulations to enforce it.

So... Dead Internet ahoy!

Reply on news.ycombinator.com