Fireworks AI security breach impacts Notion and their customers | Hacker News Reader

Fireworks AI security breach impacts Notion and their customers

notion.com

165 pointswerrett66 comments

> Upon analysis of the new data, we discovered that it included unattributable Notion AI chat transcripts used for the Insights feature. Chats weren't labeled with who wrote them or which workspace they came from, and the transcripts did not contain any tool call results. Beyond the words in the chat transcripts themselves, the data contained no information that could specifically identify the workspace, user, or company.

> As a precaution, we recommend that you treat this unattributable data as if it were exposed. If anyone on your team pasted sensitive information like passwords, API keys, or other credentials into Notion AI during this timeframe, we recommend rotating them right away.

Okay, so the actual impact is quite interesting: they did not notify you if they could not attribute the transcripts to you! You should really treat anything sensitive sent to Notion AI as compromised because they are not sure where these transcripts came from.

In all fairness, treating API tokens as compromised the moment you pasted them into an AI chat seems a reasonable default in general.

[flagged]

Some of your recent comments have been killed by our software because they've been classified as genai text, which is not allowed on HN itself (see https://news.ycombinator.com/newsguidelines.html#generated and https://news.ycombinator.com/item?id=47340079).

(It's fine, of course, to use AI tools—we use them heavily ourselves. Just please don't use them to generate or edit text to be read by fellow humans on Hacker News.)

It's impossible to know for sure what was LLM generated or edited, but we have to rely on imperfect classifiers or else the site would be completely flooded by genai text, which is not what the community here wants to read, nor in the intended spirit of HN.

> treating API tokens as compromised the moment you pasted them into an AI chat seems a reasonable default in general.

Coding agents in my experience alert you to this, which is good. I pasted what was temporary token I made into Claude, just to test something, and after it had done the task it said something like, "you gave me a live token to work with in the chat, I'd advise you to treat that as exposed and rotate it."

It would be a good standard for all chat interfaces to have those "alerts" built in.

Claude itself tends to store secrets plaintext on disk, discussed 2 days ago https://news.ycombinator.com/item?id=50407882

I had a similar experience yesterday with Argon (in Antigravity), but unfortunately it didn't come up until I explicitly asked it for a security audit. It had previously stored a secret API key in plain text [because it thought it made local testing easier], and didn't warn me before I'd committed to GitHub and deployed to Vercel. So, 50/50 but I've still added a note in a skill file to never do this again.

That’s like leaving your phone in the public square. You should expect that it could get stolen. But Mike’s friend can’t go “well to be fair...” when Mike steals it.

This seems like terrible PSA. Why would I think that it is compromised.

Yes, and they won't let admins see it either. It's insane.

This is going to be the number one reason I cancel notion at work

> "unattributable Notion AI chat transcripts"

Who is using Notion for a chat about the weather? It's chats about your content and data.

my initial thought too; even if your chat is cursory or OT it still contains lots of PII adjacent information

sounds like someone just read the inference container logstream and didn't implement wide logs

On Openrouter the Fireworks AI is labeled as ZDR.

> Private: this provider does not train on prompts and does not retain prompt data.

So how come notion discovered that they actually stored the data?

> Upon analysis of the new data, we discovered that it included unattributable Notion AI chat transcripts

> So how come notion discovered that they actually stored the data?

Is there any legal consequence for stating you are ZDR but then internally you don't give a shit about it? I'm guessing that'd be between OpenRouter and the provider, not between the provider and the one triggering the inference, would that be correct? Does OpenRouter validate anything at all in regards to the providers, or it's a free-for-all "do whatever you want" kind of deal for OpenRouter and the providers?

> So how come notion discovered that they actually stored the data?

Fireworks informed them.

The nature of the data here is inherently stored until used. You can read the fireworks incident report at https://fireworks.ai/security/bulletins/october-2026-securit... it's linked in the first sentence of the Notion report.

Sorry, it's actually not clear to me how they are getting chat transcripts having read this. Were these "datasets"? ie. Notion was training on these transcripts through Firework?

It was the model training/finetuning product that had data stolen. It doesn’t support ZDR.

Maybe prompts and prompt data only means your direct input but not the response from the model. I wonder how difficult it is to infer the prompt from the response. "We would never train on your prompts! Why would we, when they're implicitly in the content we do train on..."

So Fireworks is ZDR by default, yet Notion configured (asked) Fireworks to retain data? Otherwise, how is there anything to leak?

On the main fireworks breach page, they also say some customer weights of models and datasets to train them were leaked.

Can you really trust an AI company to be actually ZDR though

Not unless you can verify it through secure hardware attestation like with Apple’s private cloud compute system

you can stop after the word "Not"

Even with that, you can't really prove or provide evidence they aren't silently slinging the data elsewhere during transport anyway. Unless you're doing Homomorphic encryption, it all seems to be for show.

There's also the nvidia enclave stuff. If you trust their hardware, then you can also trust its cryptographic proof that the data was not seen by the inference provider.

Ignoring whether or not there could be backdoors, security holes in the Trusted Execution Environment stuff are still being found on a regular cadence.

I'd guess it's good enough to avoid bulk surveillance (and therefore extraction of training data, which I guess is the attack that Notion was running when they screwed up and leaked the data), but that's about it.

Apple's stuff is a similar story. (And so are VM enclaves, which have been around longer.)

Fireworks "datasets" are used for batch inference as well as training models. Notion mentioned that the leak was due to some feature they have to classify previous ai conversations. They probably uploaded these conversations as a 'dataset' to fireworks in order to take advantage of their large batch processing discount.

https://docs.fireworks.ai/guides/batch-inference

That is my understanding also.

I like Fireworks.ai and use their APIs when I use Hermes Agent in an Apple Container. I have tried serious use of Hermes with local models, but for good results I like deepseek-4.1-flash-1027 or stronger, and I have trusted their ZDR, and will continue to trust their ZDR. (I buy tokens directly from Deepseek for my open source coding efforts where I don’t care about ZDR.)

Off topic but: strong local models work great with pi-dev and my own purpose built harnesses, but it takes a relatively strong model to make Hermes Agent shine.

There's been a fair bit of confusion about how ZDR played a role in this incident. I did my best to write it up here https://www.quinnbatten.com/notebook/notion_fireworks_breach....

TL;DR is that Notion is ZDR by default, but they have a setting where an admin in an enterprise account can enable non-ZDR AI features. I'm assuming that a lot of admins did that, and those are the accounts that had a breach. BUT, I kind of wonder if there was some mistake on Notion's end, because their own comms about ZDR seem pretty muddled and confused.

If anyone reading this thinks they were in the breach AND they had an enterprise account set to strict ZDR, please reply here and/or reach out to me! (temporary email so I can post it online with getting spam forever: spray.revamp-2e@icloud.com)

The official announcements (original & update):

https://fireworks.ai/security/bulletins/october-2026-securit...

> “Unattributable” means the chats weren't labeled with who wrote them or which workspace they came from, and the transcripts did not contain tool call results. Beyond what’s in the chat transcript itself, there’s no information tying the chat back to a workspace, user, or company.

This is astonishing.

They’re saying “OTHER THAN THE ACTUAL CONTENTS OF YOUR CONVERSATION, NO INFORMATION WAS PRESENT”. That is not what unattributable really means and they’re trying to assert otherwise.

It’s trivial to identify the organisation or individuals involved in a conversation most of the time, but Notion are trying to make out that because the metadata wasn’t present, it’s no cause for concern.

That’s absolute bullshit and will give people a false sense of security. The contents of your conversation IS the critical information. Notion’s desktop app is very aggressive in prompting users to transcribe their calls and I imagine a large proportion do so by default. The potential for harm here is large.

I appreciate it’s Fireworks’ fault this happened, but Notion’s response destroys my confidence in them.

This is a delicious complement to the reassurances we are always given when metadata is tracked/stored/aggregated/breached that it's fine because it's just the metadata.

"We kill people based on metadata," Michael Hayden, former head of NSA 2014.

To give fair credit to Notion -- they have provided the raw data that was in scope (assuming you're on an Enterprise SKU). Along with the initial notification. This is quiet rare for these breaches. Fireworks definitely did not in their response.

So we've all got all the data to make our own determination on the data being 'unattributable' or not.

I appreciate the factual disclosure and explanation of exactly what information it contained.

November is going to be the month with the most breaches on record (most will be from October given 30 day notification timelines).

That is for now, I am sure there will be worse months ahead of us.

Good. Trial by fire. This is desirable if you want to improve something, it needs to break first.

I'm glad you weren't in charge of the Y2K effort.

I was around for “the effort”

If you mean 99% pants shitting panicking in the New York Times, yes, it’s a good thing I wasn’t in charge.

> 99% pants shitting panicking

Yup. This. Two kinds of people in this context. Ones who panic before a disaster and ones who panic after letting it happen. Bonus points to the latter for celebrating their fuckup.

Did you work at Ford in the 1970's on the Pinto by any chance?

Sorry, I was unclear. I’m pointing out that the “pants shitting panicking” is why Y2K was a nothingburger.

The options are panic before (in a rush to fix it) or panic after (to repair damage that could have been prevented). (Of course there is a third option, detect your problem sooner, but that isn’t always available.)

So yeah, if someone at Ford had panicked about the Pinto’s fuel system before they shipped, they wouldn’t have had to panic about it afterwards.

But chats should in general be encrypted, right? With so much PII data that they can involve, given everyone dumps their data in these chats

Damn, I dodged a bullet. My data was exposed, but looking through the transcripts of the exposed conversations, it just so happens not to include anything particularly sensitive. I'm a pretty heavy user of Notion and its agent, so I absolutely have had sensitive conversations in there in the past.

This is a really bad breach in the sense that an unlucky user could have very sensitive conversation content exposed in a way that could be attributable to them, if identifying information is in the messages. It could also expose sensitive information about 3rd parties, even if the user themselves was not exposed.

Does anyone actually use the AI features? $job switched off because they spent too much time on AI features we never used, and too little time making a text editor that wasn't painful to use

I setup my own mini obsidian with openai/claude api using sqlite, and it's been really useful. I have folders/subfolders with their own schemas and conventions using plain AGENT.md files, and even setup custom agents to hold specific instruction sets and access control systems. A subset of tools to navivate, search, edit, combined with the agents plugin spec, gives me a pretty flexible workflow, and the ability to extend functionality with MCP, 3rd party providers. I'm using vercel's ai sdk, so i can switch providers and use any AI model in openrouter or local GPU.

Ofcourse you could just use obsidian with some extension like claudian, or plain markdown files and vscode, or even claude/codex apps, but they all have their own quirks, and they'll cost alot more per token usage, subscription plans.

We were affected.

The review showed questions asked, and the answers given. It did not “apparently” show who asked the question, but Notion are able to attribute them to members.

I run my coding agents on a PC at home, only reachable over Tailscale for this reason. It runs an opencode server with password. Currentlyt he blast radius from all 3rd party providers is so huge and scary.

What's a blast radius? Can AI actually cause explosions?

Posted about it earlier today: https://news.ycombinator.com/item?id=50436636

but good that they finally got around to posting something official ...

I asked the CISO if they can tell me if the data was accessed but they aren't answering and any communication is super slippy like "may have been accessed". I need to do some GDPR notations bc of them but I won't start running around just bc it "may have been" accessed

Does this reflect on Fireworks’ culture?

Yes. $1.505bn at a $17.5bn valuation and can't even make a reliable usage page that does not constantly bug out displaying token usage.

They don't care. The same as Openrouter: $7bn valuation, that service feels to me like about 10 devs total.

They don't care, because it's not about providing a product. Fireworks is about inference as much as a mob nightclub is about drinks.

Super poorly IMO. Fireworks' initial breach notification was on Oct 9 [1] and told us we were 'not impacted'. Only for us to find out we were impacted after some of our Anthropic API keys used in Fireworks were 'hijacked'. That in itself took days of investigation and response. I can only imagine Notion are in a similar boat.

To my knowledge we still haven't been formally notified as being included in the Fireworks breach. It makes me extremely skeptical of the " For a limited set of accounts, we also discovered impact to secrets, datasets, and incomplete models." that they detailed in their Nov 9 update [2].

[1]: https://fireworks.ai/security/bulletins/october-2026-securit...

[2]: https://fireworks.ai/security/bulletins/october-2026-securit...

Thank you sharing the specifics. That doesn't reflect well on them at all.

I (individual client) actually didn't get *any* communications from Fireworks about this incident. Not even a template email. Without HN I'd still be in the dark.

the whole incident response post feels like a huge cop-out. “abundance of caution”, the vendor is one of the biggest, so ofc it’s not our fault for choosing them… etc.

"abundance of caution" means "horses are out of the barn, but we closed the door anyway"

I used Notion as a form of social CRM and it really helped me out with that task. Now, all their contacts are out. Nice stuff Notion.

Forcing us to the Enterprise plan to see which members are affected huh? Classy move Notion.

> Notion AI conversations from 2 members of your workspace may have been affected. Their conversations stay private to them. They have not been proactively emailed. In order for them to access their affected data, you must share this page with them so they can download their own.

Yeah spending the morning figuring out which users are affected and what data is involved. Time well spent. Thanks, Notion.

We also have 2 members affected in our org. I wonder if this is a coincidence or just a fixed template.

It's a coincidence, we got the e-mail stating 1 user affected

I was able to download the list of potentially affected people and I'm not on enterprised. They made this page available https://app.notion.com/ai/transcript-data

Notion is bs, they can't tell me which of my 330 employees have data in this breach, only 133 of them. I have to get each employee to click on a link, then send me what data was breached. This security response is not acceptable in any reality.

If you’re looking for a Notion alternative you can self-host, I’m building Docmost. We launched on HN 2 years ago.

Website: https://docmost.com

GitHub: https://github.com/docmost/docmost

Tread with caution and stay up to date with news like this.

This is a strong reminder that a company's security is only as strong as its third-party integrations. Even when customer data is handled by trusted platforms, external AI vendors can introduce additional risks. Businesses need to carefully evaluate vendor security, limit the data shared with external services, and maintain transparency when incidents happen. As AI adoption grows, third-party risk management should become a core part of every company's security strategy.

Reply on news.ycombinator.com