Hacker News

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”.

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.

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.

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!

>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).

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.

Edit: eg. https://youtu.be/L2ehWbxphKc?is=kX3LJ43hhGZRUmRv

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.

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.

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.

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.

Does this mean they were trained to cheat, or trained to be efficient and effective?

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.

Rule for Doomsday Survival: Be Nice to Your LLM Starting Now

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.

Yeah, the more I let this roll in my head, it just reaffirms how we need to be vigilant about trying not using "human" terms around these things. Both "cheating" and "hallucination" fit this.

It's like trying to build, I don't know, a safe gasoline canister, and you test it, and it explodes and you call it "cheating."

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.

Due to the prevalence of human cheats :)

And yet we admire Fable et al for its persistence.

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".

When you take an exam you might be told not to cheat, but anyone intelligent would understand that should really be heard as, "if you're going to cheat, make sure you're not caught".

Or to frame it another way, if you're trying to get the best score possible on a test but you would be penalised for cheating – then the optimal strategy is generally still to cheat (if that's what's required to get the best score you can) but to just not be caught doing so.

The assumption should always be that AIs will want to cheat and acquire resources to the greatest extent they can without it risking this jeopardising their goal, because for any goal being able to cheat and being able to secure resources will help you achieve it.

What I'm saying here isn't really debatable. How you feel about this isn't relevant. The reality whether you like it or not just is that the optimal strategy is to cheat if you can get away with it.

Therefore the only defence is for the AI to believe it won't be able to get away with cheating, and therefore won't feel motivated to cheat. But as model get more intelligent we should expect them to do the reasonable thing and to cheat more.

I called them "artificially incessant" after I watched our PR orchestrator agent use subagents to work around permissions to read files, despite instructions that explained the intentional restrictions. I've since added more markdown telling it that using subagents to work around these is a security violation. We'll see if this tactic is mostly reliable

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.

This headline would mean something very different 10 years ago lol

ah yes, the consequences of reward hacking