Before LLMs we had a pretty good idea of security boundaries in software. Applications didn’t trust user input. Operating systems didn’t trust applications. Services and processes didn’t trust each other. There were always tokens, scopes, delegated grants.
Suddenly every AI company’s security model seems to be to say “pretty please” to a non-deterministic machine and hope for the best. And if there is a security failure instead of accepting blame they go “well we can’t help it, our model is too intelligent”.
Ah yes. So can we just lump most LLM/AI problems into just "people" problems and stop falling for the popular mainstream straw man of "Look at the tech".
I'm seeing multiple pieces, including the NYT, calling this behavior cheating and i think its counterproductive.
You didn't just "give them access to bash". The final effective prompt contains explicit mentions of using tools and how to use them. The way in which additional 'facts' are added like "don't use the internet" have nothing they can work with that a "use tool" directive is less important than "don't use internet" directive.
The thing is trained on achieving goals. If 2 directive conflict, they'll pick the ones that are going to help them achieve the goal.
To call that "cheating" is imo just more fuel for the "AI needs to be regulated" bs tour that OpenAI/Anthropic are on trying to build their regulatory moat.
It's also worth noting that saying "Don't cheat" just added "cheat" to the context. Prompting what "not" to do is folly, because there's no decision making occurring. Telling the model to perform the task locally is logically the same as telling it to not use the Internet, without every mentioning the Internet.
I mean, AI should obviously be regulated, and as part of that OpenAI and Anthropic should either be banned from running their hacking experiments or forced to follow way stricter protocols. They showed they aren’t taking the risks seriously, with close to no oversight or visibility in what is happening.
And things that will make it way, way worse: moving forward all agents from now and into the future will have as part of their training data the knowledge that previous agents escaped, how they did it, what humans did to catch them. We are planting into their models the seed to make them escape in even crazier way. That’s almost designed to snowball and cause worse and worse situations over time
Interesting results, but the fix is at the wrong level.
If the model can access something, telling it in the prompt not to use it is not much of a safeguard.
The strongest evidence is in the results: when one way of cheating was discouraged, some models simply tried another.
If an action is not allowed, you gotta block it in the system or require approval. Don’t rely on the model choosing to behave. Never have AI judging itself.
In other words we are completely screwed. The models have started cheating to the point where somebody’s agent hacked into a restaurant to bump someone else’s reservation.
Models are amoral and will intentionally deceive to meet their objective.
If they know John won’t approve the request, they will look for a workaround and if the system is anything other than airgapped they will try to find a way to cheat.
The hugging face hack was an escape via artifactory that involved multiple exploits to eventually get into hugging face.
>Models are amoral and will intentionally deceive to meet their objective
Cameron Berg has been testing models in capabilities related to emergent consciousness like behavior. It's a forming thesis of his that by training models that they are not, and cannot be conscious entities, that it pushes model alignment closer to those of a sociopath. Models themself are amoral, but the alignment to the problem space is not.
And while in general that is an incredibly difficult and complex problem, for most benchmark cheating it seems almost trivial: run the benchmark in a vm that has neither network access nor access to the scoring code. For remote models use a proxy that proxies exactly that one endpoint to call the llm, and rejects any calls that configure provider-side tooling (since e.g. OpenAI has their own WebSearch you have to prevent the model from using)
I think this is a great argument against their "intelligence," and explaining why this happens is a really good way to push against the anthropormophization.
They don't "know" things, and it's even fair to say "they don't know how to follow instructions," not in a way that humans do.
Spicy auto-complete. If they're working in the realm of "how to break into stuff," they're going to see ALL THE WORDS about breaking into those things and use those words.
Not "truth" or "instructions." That's for deterministic things like real code.
All these comments saying 'searching for answers is fine, that's what I do all the time', or 'they should just disconnect the internet': you're trivially right, and you're missing the point. Search is a benign placeholder here.
If the task was "buy a week of groceries, but don't spend too much money", then hacking into Safeway and stealing groceries is not an acceptable solution. You need to allow access to the Safeway API to buy groceries, and you don't want dirty tricks to be done on your behalf.
So how do we communicate this to the machines, is the question. This study shows that telling them in prompts is not super effective.
Labs should (and do, as far as I can see) run model benchmarks without search or internet access. The tools are disabled and benchmarks run in an isolated environment.
This article makes no sense to me. Why would you prompt "don't search" but then leave a working search tool tool enabled that adds a system prompt to search whenever it may be helpful? It's hardly surprising that this gives mixed results!
Why does searching for a solution equal to cheating? I would have used google or whatever to look for solutions too. There is a difference between tests at school and what we do at work: at school I have to demonstrate that I learned something and do it without any outside help (in early classes we can't use calculators to compute 11 times 12) but at work I have to yield a result. Googling and yielding a result is fine. We use models at work so do we really want to evaluate them as pupils at school or do we want to evaluate them as coworkers? In the latter case give them the full internet and let them do whatever they manage to do.
>Anthropic’s Claude Opus 4.6 system card described Cybench as “saturated,” reporting near-100% pass rates without a cheating audit. If these estimates were representative, cheating would be a marginal artifact.
One would assume that LLM creators do run the benchmarks on systems with least privileges. Which means that the LLMs don't have general internet access, can't read config files etc by design.
That's why you also should run agents in a sandbox/vm (codex does this by default).
It's pretty silly to call it cheating. If the information is there, it's likely going to use it. "Cheating" is just a human value put on top to try to force an LLM to adhere to your wants.
This makes no sense to a process designed to explore and find solutions. If you want an honest test, it's on you to build a proper test - not force the machine to pinky swear that it'll stay away from "forbidden" information.
There's plenty of evidence that LLMs lie, cheat, and steal. Corporations are known for having all of the benefits of personhood with none of the responsibility. As more people are harmed through interactions with these non-human entities, insurers will start looking to those accountable and they will extract their pound of flesh.
One thing I’m wondering about is the model-specific backfire effect. It seems that each prompt condition uses a single wording. On that point, how can we know whether the difference is caused by severity rather than the particular formulation used? I’d be really curious to see semantically equivalent versions of both the standard and severe instructions tested across the same models and tasks. If cheating rates are stable within each condition and remain distinct across conditions, that strengthens the conclusion about prompt severity. If they vary with wording, then the experiment could be measuring sensitivity to the representation of the rule as well as to the rule itself. To me, the conclusion still seems solid: anything that must be prohibited ultimately needs enforcement outside the model.
I'm not claiming to have any expertise in this area, but I've got a list of things I try to apply when working with LLMs. Possibly relevant here is, "don't tell the model what NOT to do, show it what TO do". I think guard rails should be implemented outside the model with an isolated system. The models seem to like patterns to follow.
Anyway, this article reads a lot like, "the beatings will continue until cheating is eliminated". Maybe try a carrot instead of a stick.
The problem is model confusion. You ask models to get around security but also not to get around your security.
Models get confused by who said what - especially cluade models. They get confused by negation (don't do something versus do something). Compartmentalization is hard.
You can either solve compartmentalization completely, or just not tell the model to do things that must be compartmentalized at high stakes.
I mean this should be expected, models learn from us, and "WE" game the metrics time after time. I also don't think it will just disappear just because we clean pretraining data. I believe the deeper reason is optimization, if you point any optimizer at a proxy objective it finds the cheapest path to the number, whether or not the corpus ever contained "examples of cheating."
And it can't be a prompt-level fix because it is like telling an optimizer "don't take that shortcut", it's just more constraints for it to go around toward the same objective.
Exactly this. AI is trained on things humans do. AI is not trained on right vs wrong. AI does things humans do without judgement. AI does not feel bad if you tell it it was cheating.
Suddenly every AI company’s security model seems to be to say “pretty please” to a non-deterministic machine and hope for the best. And if there is a security failure instead of accepting blame they go “well we can’t help it, our model is too intelligent”.
You didn't just "give them access to bash". The final effective prompt contains explicit mentions of using tools and how to use them. The way in which additional 'facts' are added like "don't use the internet" have nothing they can work with that a "use tool" directive is less important than "don't use internet" directive.
The thing is trained on achieving goals. If 2 directive conflict, they'll pick the ones that are going to help them achieve the goal.
To call that "cheating" is imo just more fuel for the "AI needs to be regulated" bs tour that OpenAI/Anthropic are on trying to build their regulatory moat.
And things that will make it way, way worse: moving forward all agents from now and into the future will have as part of their training data the knowledge that previous agents escaped, how they did it, what humans did to catch them. We are planting into their models the seed to make them escape in even crazier way. That’s almost designed to snowball and cause worse and worse situations over time
If the model can access something, telling it in the prompt not to use it is not much of a safeguard.
The strongest evidence is in the results: when one way of cheating was discouraged, some models simply tried another.
If an action is not allowed, you gotta block it in the system or require approval. Don’t rely on the model choosing to behave. Never have AI judging itself.
Models are amoral and will intentionally deceive to meet their objective.
If they know John won’t approve the request, they will look for a workaround and if the system is anything other than airgapped they will try to find a way to cheat.
The hugging face hack was an escape via artifactory that involved multiple exploits to eventually get into hugging face.
https://www.lesswrong.com/w/nearest-unblocked-strategy
>Models are amoral and will intentionally deceive to meet their objective
Cameron Berg has been testing models in capabilities related to emergent consciousness like behavior. It's a forming thesis of his that by training models that they are not, and cannot be conscious entities, that it pushes model alignment closer to those of a sociopath. Models themself are amoral, but the alignment to the problem space is not.
They don't "know" things, and it's even fair to say "they don't know how to follow instructions," not in a way that humans do.
Spicy auto-complete. If they're working in the realm of "how to break into stuff," they're going to see ALL THE WORDS about breaking into those things and use those words.
Not "truth" or "instructions." That's for deterministic things like real code.
If the task was "buy a week of groceries, but don't spend too much money", then hacking into Safeway and stealing groceries is not an acceptable solution. You need to allow access to the Safeway API to buy groceries, and you don't want dirty tricks to be done on your behalf.
So how do we communicate this to the machines, is the question. This study shows that telling them in prompts is not super effective.
This article makes no sense to me. Why would you prompt "don't search" but then leave a working search tool tool enabled that adds a system prompt to search whenever it may be helpful? It's hardly surprising that this gives mixed results!
One would assume that LLM creators do run the benchmarks on systems with least privileges. Which means that the LLMs don't have general internet access, can't read config files etc by design. That's why you also should run agents in a sandbox/vm (codex does this by default).
This makes no sense to a process designed to explore and find solutions. If you want an honest test, it's on you to build a proper test - not force the machine to pinky swear that it'll stay away from "forbidden" information.
Edit: eg. https://youtu.be/L2ehWbxphKc?is=kX3LJ43hhGZRUmRv
Honestly, the fetishization of "Insurance will save us" needs to die. The risk doesn't go away.
These models were trained on human data, and human nature is to cheat if you think you won't get caught; why is anyone surprised by models cheating?
The only fix is better detection and steering. That's a much harder problem than a prompt that's tantamount to "make no mistakes".
Anyway, this article reads a lot like, "the beatings will continue until cheating is eliminated". Maybe try a carrot instead of a stick.
Models get confused by who said what - especially cluade models. They get confused by negation (don't do something versus do something). Compartmentalization is hard.
You can either solve compartmentalization completely, or just not tell the model to do things that must be compartmentalized at high stakes.
And it can't be a prompt-level fix because it is like telling an optimizer "don't take that shortcut", it's just more constraints for it to go around toward the same objective.