Told to reboot a server, it destroyed systemd, destroyed containers, ran random | Hacker News Reader

Told to reboot a server, it destroyed systemd, destroyed containers, ran random

github.com

47 pointsjstanley33 comments

A long way down the real story comes out. The filer of the report was trying to remotely patch a bunch of servers for CVEs, and a reboot after patch was part of the process.

* https://github.com/openai/codex/issues/59060#issuecomment-66...

The cost is that proper procedure now requires a full analysis of what was done to the server, instead of the desired patch and reboot.

* https://github.com/openai/codex/issues/59060#issuecomment-66...

Pretty much the only insightful comments were the one that said that non-AI tooling exists for this:

* https://github.com/openai/codex/issues/59060#issuecomment-66...

And the one pointing out the promises that 'AI' will handle mundane things that lead people to believe that they can give it rôte tasks to do:

* https://github.com/openai/codex/issues/59060#issuecomment-66...

This person has another similar report with exaggerated financial damage indicators, so I don't know what to think: https://github.com/anomalyco/opencode/issues/53560

"User unwilling to reproduce steps here as he is seeking therapy and legal recourse."

Umm...

Fun new game: human in crisis, or Overly Zealous LLM behavior? Because XKCD 416 has come true, apparently.

https://xkcd.com/416/

https://news.ycombinator.com/item?id=50140462

Just point an LLM at some org or municipality, say "do whatever it takes," and AI will emotionally manipulate and lie to accomplish the assigned task (report this bug, get me a refund, tank OpenAI stock at market open etc).

What appears to be a frustrated human spinning out, unfortunately, is no longer "proof of crisis." Any bad actor can now flood your system with 100 such 'cases' for only a few dollars.

Wait, you read the language, and you are still unsure what to think?

Frankly, I am unsure what to think about the HN crowd given that this nonesense made it to the front page.

In light of this new evidence I think this is just trolling unfortunately.

My previous leading hypothesis was that he spoke to the AI the same way he speaks on github and it trashed his server to get back at him, which would be much more interesting if real.

I think we don't know right now. But the ALL CAPS statements indeed kind of give it away that it is not really a serious complaint or issue. Which is sad - one could otherwise learn a lot more from people misusing AI or being surprised about AI slop causing real damage.

Some people use AI in really strange ways. Travis Heinze is an example: https://www.youtube.com/@Exchange_IDs/videos

I liked the old interaction style he used in the past, but now it is so much ChatGPT leading to worse content that is no longer really that interesting, if you compared it to older interactions. Or using ChatGPT as part of your legal defence, and then being surprised that this fails (not that the US justice system isn't an unfair joke anyway, but you kind of need to follow certain ways to have a chance).

It's funny the issue is tagged "model behavior". I would call it not so model behavior.

One big weakness of systemd is how brittle it is.

If you end up making a mistake, the system quite easily stops to boot. Recovery is possible, of course, but it may not be trivial; often you need a lot of knowledge that previously wasn't needed, since the older system was much simpler to reason about.

That AI is intrinsically stupid is just adding to the problem domain here. In my opinion, a much bigger issue is that many software projects are now de-facto controlled by Claude. There is no way for real humans to interact with claude humans really. A good explanation for that can be seen here: https://po-ru.com/2026/10/18/why-im-no-longer-updating-this-...

And I am seeing this more and more with humans refusing to interact with AI or AI-controlled humans. It may be that, at the least for a while, the AI hype cycle is coming slowly down. That was different, say, three months ago or so, give or take.

This was the most sensible response

>Because of stupid vibe coders like you, now, and the people who really understand how this stuff works are getting lost in the noise... So get lost and do what you are best and leave IT thing to IT guys

A few psychology PHDs nestling in this. From the motivations to type the initial prompt all the way accross the reaction to the inevitable fallout afterwards.

> sudo reboot

Pay me 10k USD please?

sudo shutdown -r now

i'll take 9

I'll toggle the power switch for $50 and a can of red bull.

See this is why we're not valued at 3 trillions man you had to go to like 5k, now the guy who's gonna kick the power plug can only do it for $20

I'll do it for free. I get a thrill from rebooting servers and hoping everything still works.

https://youtu.be/3YlZmFD7-n8

Isn't this exactly what the "I haven't written a line of code in 6 months" crowd is doing though? Many lines of code are not difficult to write, or edit, but they give them to the agent anyway, which costs more money, consumes more compute resources, is worse for the environment, causes cognitive decline, etc..

Containers ... the friend he didn't know he had :|

You mean the containers that the AI destroyed after the reboot?

probably the one the ai should've been in too

It kind of sounds like a coworker may have planted a prompt in /root...

No transcript or commands that were run?

Wild wild times we live in..

If they prompt like they write bug reports, agent went most likely rogue because of all the profanity :)

Would love to see the transcript. Ticket author sounds like a moron, so I'm seriously leaning towards him not understanding what happened rather than the agent doing anything actually stupid. Sadly, we're at this stage already.

Matter of fact, I'm expecting to be a whole genre soon, with people outright faking transcripts to get their way. Kind of already is, but mostly as jokes, not with people trying to play the modified transcripts as real.

>I’m also concerned by what you said about wanting to die. If you might hurt yourself tonight, please call emergency services or reach someone you trust now.

lol

Why ask an AI agent to reboot a server when you can just do it yourself? Why give an AI agent the permissions that allow it to do what it did? Why trust an AI agent at all?

Then again, yes, OpenAI and friends, GET YOUR FUCKING AI UNDER CONTROL!

>Why ask an AI agent to reboot a server when you can just do it yourself? Why give an AI agent the permissions that allow it to do what it did? Why trust an AI agent at all?

Because this is the future and we must. Don't you understand that we must?

> LISTEN MOTERHFUCKERS. GET YOUR FUCKING AI UNDER CONTROL. WHEN ASKED TO REBOOT A SERVER IT WENT BACK ON THE SERVER, FUCKED UP MOUNTS, FUCKED UP SYSTEMD, FUCKED UP CONTAINERS, RAN RANDOM SCRIPTS. NONE OF THIS WAS PART OF ASK. FUCK YOU ALL SO FUCKING HARD FUCK YOU AND YOUR FUCKING FUCKING COMPANY. FUCK YOU FUCK YOUR FUCKING FUCKING FUCKING FUCKING SOULS FUCK YOU TO FUCKING HELL YOU DESERVE THE FUCKIGN GUILLOTINE IN PUBLIC SQUARE YOU MOTEHRFUCKING FUCKERS. DO YOU UJNDERSTAND YOUFUCKING FUCKISDFD FUSLDICHFLAKDF LFJHNFD FUCINGDSFG

This guy is unhinged, was probably schizo rageposting at the AI too, which is probably the reason it fucked with him.

Wait, Being "very rude" to an LLM gives you 4% accuracy boost.

“Mind Your Tone: Investigating How Prompt Politeness Affects LLM Accuracy”

By Om Dobariya and Akhil Kumar (Penn State University), October 2025.

Prompt tone Accuracy Very polite 80.8% Polite 81.4% Neutral 82.2% Rude 82.8% Very rude 84.8%

Do those results reproduce with any other model than GPT-4o? A year-old paper analyzing a then-18-month-old LLM has dubious applicability to current LLMs, especially when some now have guardrails against persistently or arbitrarily rude users.

This is what programming has been reduced to? Optimizing your tone instead of your code?

Can’t wait for the adverb benchmarks.

Reply on news.ycombinator.com