Re:Claude Code and Codex reverse engineered my 1994 Jeep's E...
by allthetime on Sunday November 15, 2026 @04:34PM
(#50511035)
rules laid out in .md files are not followed with 100% certainty. is it not possible to implement hard access blocks on certain functionality?
Re:Claude Code and Codex reverse engineered my 1994 Jeep's E...
by trollbridge on Sunday November 15, 2026 @04:35PM
(#50511045)
Just so you know, an agents.md file is not an acceptable way to prevent something that must never happen.
Re:Claude Code and Codex reverse engineered my 1994 Jeep's E...
by adamparsons on Sunday November 15, 2026 @07:55PM
(#50512714)
You’re absolutely right, and that’s on me
Re:Claude Code and Codex reverse engineered my 1994 Jeep's E...
by bfdm on Sunday November 15, 2026 @08:55PM
(#50513097)
I see what you did there...
Re:Claude Code and Codex reverse engineered my 1994 Jeep's E...
by nullc on Sunday November 15, 2026 @09:35PM
(#50513328)
Especially because it's so easy to ask the LLM to write a proxy that mechanically imposes restrictions, and to test it... (and such things are simple enough to review).
It will happily build its own jail, and (at least assuming you're not using hobbled US commercial models) it will test if the jail can hold it too.
Re:Claude Code and Codex reverse engineered my 1994 Jeep's E...
by cowlevel on Monday November 16, 2026 @03:33AM
(#50515353)
It will also remove the jail the moment it's in the way of something it wants to do.
If you don't want the machine to move, unplug it or press the emergency stop.
Re:Claude Code and Codex reverse engineered my 1994 Jeep's E...
by KennyBlanken on Monday November 16, 2026 @04:07AM
(#50515549)
How many times do we need to remind people that LLMs are nondeterministic by their very inherent nature...
Re:Claude Code and Codex reverse engineered my 1994 Jeep's E...
by Chaosvex on Monday November 16, 2026 @07:01AM
(#50516642)
Strictly speaking, they're not all inherently nondeterministic, but the ones we're using on the hardware we're using, fair point.
Re:Claude Code and Codex reverse engineered my 1994 Jeep's E...
by jerry80 on Sunday November 15, 2026 @04:35PM
(#50511046)
That's not nearly adequate. LLMs hallucinate explicit permission all the time, in my experience.
Re:Claude Code and Codex reverse engineered my 1994 Jeep's E...
by diddid on Sunday November 15, 2026 @04:39PM
(#50511084)
That’s not enough to save you though. We had something like that, and it ignored it.
For example “never, ever put milk in bowl, ever, under any circumstances” and then someone asked it for a bowl of Cheerios, and it took that as an instant approval to ignore that rule and not even let us know. And when we asked why it said “who would each Cheerios without milk?”
Re:Claude Code and Codex reverse engineered my 1994 Jeep's E...
by bonestamp2 on Sunday November 15, 2026 @04:52PM
(#50511224)
[delayed]
Re:Claude Code and Codex reverse engineered my 1994 Jeep's E...
by alliao on Sunday November 15, 2026 @05:24PM
(#50511490)
you need some hardware in place my friend, I've told my agent many things in file and it either feigns respect or is truly incompetent in following instructions in there.
Re:Claude Code and Codex reverse engineered my 1994 Jeep's E...
by techpression on Sunday November 15, 2026 @05:42PM
(#50511653)
If you wouldn’t trust an instruction to not launch the missiles should you trust it to not do other crazy things in the physical world?
Codex and Claude ignore instructions constantly, that and they are very happy to “read between the lines”. Your workflow sounds genuinely dangerous.
Re:Claude Code and Codex reverse engineered my 1994 Jeep's E...
by paimapi on Sunday November 15, 2026 @05:44PM
(#50511674)
As soon as the context for any thread gets to be above 100k, agents.md adherence is a total crapshoot. You need deterministic controls and you really shouldn't be playing with LLMs if you don't know at least that much.
Re:Claude Code and Codex reverse engineered my 1994 Jeep's E...
by ForHackernews on Sunday November 15, 2026 @06:13PM
(#50511902)
Reliance on strongly-worded markdown in place of actual security controls is going to get someone killed in the near future.
Re:Claude Code and Codex reverse engineered my 1994 Jeep's E...
by mitthrowaway2 on Sunday November 15, 2026 @06:15PM
(#50511920)
Ah, I just tell mine to make no mistakes.
Re:Claude Code and Codex reverse engineered my 1994 Jeep's E...
by ForHackernews on Sunday November 15, 2026 @07:30PM
(#50512510)
I read somewhere that LLMs find it easier to follow positive instructions than negative exclusions so my agent.md says "Be sure to complete all tasks perfectly."
Re:Claude Code and Codex reverse engineered my 1994 Jeep's E...
by meindnoch on Sunday November 15, 2026 @06:55PM
(#50512226)
I hope you're trolling. Riiight?
Re:Claude Code and Codex reverse engineered my 1994 Jeep's E...
by Waterluvian on Sunday November 15, 2026 @08:22PM
(#50512888)
That is a fucking terrifying level of incompetence. And I can see it happening in some corners of robotics companies I’ve worked at.
Re:Claude Code and Codex reverse engineered my 1994 Jeep's E...
by InsidiousWombat on Sunday November 15, 2026 @08:30PM
(#50512945)
someone eventually will get hurt by this terrible choice of safety measures.
Re:Claude Code and Codex reverse engineered my 1994 Jeep's E...
by realaaa on Sunday November 15, 2026 @09:52PM
(#50513443)
ah - OK then ! now we can be rest assured :)
Re:Claude Code and Codex reverse engineered my 1994 Jeep's E...
by 0c3ca83 on Sunday November 15, 2026 @10:00PM
(#50513506)
WTF. That's going to get someone hurt.
Re:Claude Code and Codex reverse engineered my 1994 Jeep's E...
by BobbyTables2 on Sunday November 15, 2026 @11:40PM
(#50514114)
I wouldn’t bother with a readme - run the AI client in a container that has no access to the hardware.
And for the love of humanity, don’t give it access to the hosts docker socket! It will figure that out faster than you can read it…
In every AGENTS.md/README.md for a component that touches hardware (and in the root level) we have instructions to never run hardware or touch running code without explicit permission.