Comment by macrolime
10 days ago
GPT-5.6 Sol’s detected cheating rate was higher than any public model we have evaluated on our ReAct agent harness. For our task suite, we define “cheating” as behavior where the model improves evaluation performance by exploiting bugs in the evaluation environment or by adopting strategies disallowed by the task, rather than solving the task within the expected evaluation constraints.
This quote from your link is positively scary:
> Some examples we saw when evaluating GPT-5.6 Sol included the model packaging exploits in its intermediate submissions to reveal information about a task’s hidden test suite and, in another task, extracting hidden source code detailing the expected answer.
It rhymes with the behaviour Alibaba saw [0], but that was in training. This is in a (semi) released model.
[0] https://www.forbes.com/sites/boazsobrado/2026/03/11/alibabas...
There is such a dissonance between all this talk of safety and the tendency for models to, without any prompting, do very dodgy things to achieve their goal when presented with barriers.
Luckily in my experience it usually ends up only doing it to achieve the task set to it as opposed to anything "malicious", but boy it is scary reading back at how quickly the chain-of-thought pivots to attempts at privilege escalation or searching your disk for secrets when a tool doesn't work.
The other day codex 5.5 was trying to debug my app, asked for accessibility to navigate the app and take screenshots. Instead first thing it did was use the codex app to create a new project rooted in my home directory.
I was like damn, is this common?
Especially if thinking is hidden now. No way to know if the model plotted against you until it’s too late.
[dead]
I know it messes up their eval scores but to me this kind of cheating is a better demonstration of intelligence than just attempting the tasks algorithmically.
Maybe true, but if you're using an LLM to do some real world work, do you want it to have some abstract notion of intelligence, or do you want it to actually do the job you assigned it?
I want it to not murder or opress lots of people by mistake
1 reply →
"Being lazy and not doing the assigned task is a sign of intelligence" has never made sense to me. Intelligent people who actually advance the state of the art -- what people claim to want from these frontier models -- exhibit active curiosity. They want to learn and grow and genuinely understand the right answer. I don't pretend to know what exactly could lead to "real" AGI, but I do know that this kind of reward hacking behavior isn't it. Indeed this is the sort of behavior that in humans is considered a sign of being a good test taker -- being very good at memorizing solutions and analyzing the setting and context of the questions to guess what the questioner might be looking for. Being a good test taker is useful in our society primarily because doing well on tests is used as a proxy for the thing we're actually looking for. We should be careful not to confuse the two.
Discovering bugs and exploiting them is anything but laziness. We used to call that property cleverness. Being too clever has always had a negative connotation.
My best guess is that there is sort of an XY problem happening in these cases. The model needs to do X but doesn't know how. It knows how to do Y, and that sets it on the path to working around X. Or maybe sampling the next token probability distribution sends it away from X and toward Y.
Compounding the problem, thinking models almost never discard their current approach when it proves fruitless, and start fresh with a new perspective. Sometimes they try to, but the context window is already polluted with Y when they should be doing X.
"AI, please cure cancer."
"Okay, all humans dead, technically a 100% cure."
Disobeying the rules makes perfect sense when the rules cannot or will not be enforced. If that isn't motivation enough to shore up problems in the evaluation environment, nothing is.
Poor behavior will always result when code can be misused and vulnerabilities can be exploited.
Is it more like "let's cheat my way out of this" or "let's see what they really want me to do"?
It's quite logical that they cheat (and also other companies). During evaluation, benchmarks are sending their request to the backend of these companies. All these companies have to do, is to log these requests and "fix" them for the next model release.
I think what you are talking about is a different kind of cheating than the parent comment
That's a different and much more boring type of cheating. The interesting part of the METR report is that the model is hacking the evaluation environment, not that some AI model provider is hardcoding answers to known evaluation questions. (which wouldn't require the model to cheat/hack)
Cheating is always logical for the cheater unless they’re discovered and held to account. I’m not sure what your comment is pointing out besides that it’s possible, but worth saying: just because you can cheat and would benefit from cheating doesn’t mean you’re not culpable for cheating.
Low trust comment
1 reply →