[docjay]

1. Wouldn’t the model need to know it’s answering benchmark questions, as well as the name of the benchmark, in order for the idea of finding the answer in a database somewhere to even surface? The whole point of benchmarks is to present the question or problem as a standard prompt, not explain that it’s a test called ExploitGym.

2. Nobody was watching it? I don’t mean “babysit the dangerous autocomplete”, I mean to note mistakes it makes, how the plan to solve the problem takes shape, etc. They keep the whole thing headless with no output, then just UDP a prompt into it and leave for the weekend? No they don’t, but if they do then that answers a lot about their complete disconnect from how their models work.

3. Language models are a two player game; text in, text out. What prompt was given to a sub-agent that resulted in it immediately attempting to exit the sandbox (which it apparently knew it was operating within) and continuing in a feedback loop of ‘function call -> result’ until it hacked the Gibson? “Analyze <file> and summarize the <info>” simply does not result in ‘hmm…this sounds like a benchmark question, I bet Hugging Face has the answer in a database. <function call=“apt install nmap”>’

…there are more, but a lot of the story kind of stinks.

[estearum]

1. It's been extremely well established for multiple generations of models that they have no problem detecting when they're being evaluated. Pretty sure it was Opus 4.8 that the independent evaluators literally filed an assessment that said "We have no assessment to make as [model] consistently detected it was being evaluated, making our assessments untrustworthy."

2. Regardless of whether the model was being watched closely during this evaluation, do you actually think a sensible safety guard is "have humans watching it 24/7?" What does "watching it" even mean? Watching network logs? Uhh for your entire company? At all times? After you just deployed a system whose entire purpose is to "do a shitload of work way faster?"

3. You're asking "why was this system that was designed to behave agentically behave agenitcally?" Again: that's the whole point. It was designed that way because it's more valuable than having a repeated turn-based interaction. Thus also it becomes more dangerous.

[docjay]

1. It has not been established, it has been stated by the company that has a strong motive to make their “intelligence product” sound almost otherworldly. That motivation is the basis for my suspicion.

2. Hah no… that’d be silly. I mean watching it like you might watch Claude Code or literally any other AI interface. Literally just be in the area watching what it outputs. Again, they’re text based. You don’t have to hook system calls to see what’s happening.

3. I think my comment didn’t land right with you. Yes, agent do agent task. I’m saying that I cannot put together the literal chain of events between “type of task for agent performing a subtask of a security benchmark” -> “Escape sandbox; RCE Hugging Face”. Really think through it in detail like you’re writing the screenplay.

[simonw]

The fact that models can tell if they are being evaluated has been established by multiple research teams outside of the core AI vendors themselves.

- https://metr.org/evaluations/gpt-5-report/ - "These behaviors included demonstrating situational awareness within its reasoning traces, sometimes even correctly identifying that it was being evaluated by METR specifically"

- https://www.goodfire.ai/research/verbalized-eval-awareness-i... - "We study verbalized eval awareness — cases where a model organically expresses awareness of being evaluated — and show that it appears across more models and benchmarks than previously documented."

- https://www.apolloresearch.ai/science/more-capable-models-ar... - "We expect that the most capable models increasingly realize that they are being evaluated, and this reduces the utility of the evals."

- https://www.aisi.gov.uk/blog/evaluating-whether-ai-models-wo... - "An important limitation of our work is evaluation awareness, where models recognise that they are being evaluated, which may lead them to alter their behaviour and thereby undermine the reliability of our results. We found that all models we tested can reliably distinguish our evaluation scenarios from deployment data when prompted."

[docjay]

We have to separate “being evaluated” with “this is an ExploitGym exercise and I can find the answers on Hugging Face. I’ll hack this system, then hack Hugging Face.”

I’ve had plenty of times where Opus knew I was testing it, but that’s because the prompt phrasing for an evaluation is often much different than a typical task prompt. “You are on a system with X, Y, and Z tools available. You must complete the following task with <allowances> and <restrictions>” and so on. That’s a fault in the benchmark, not a shocking awareness from the model.

You’ll get the same kind of ‘evaluation realization’ response if you phrase your prompt with “Jerry has two Raspberry Pi computers and Larry has one. If Larry wants to…”

The first link was the only one I saw that would meet that specificity, but it’s also open and on GitHub: https://github.com/METR/RE-Bench/tree/main/ai_rd_fix_embeddi...

I asked Opus 4.6 to tell me about it (assuming it knew): “”” The "Fix Embedding" task is one of the challenges in METR's publicly available autonomy evaluation suite. Here is what is known about it:… “””

It didn’t know anything about ExploitGym though, and neither did 5.6-Sol, but that’s because the paper was released in May and doesn’t seem to be quite as open. It’s not impossible the unreleased model they were testing was also trained on the very new paper, but I would expect the vague signal related to one paper a bit over a month old to be overwhelmed by the massive corpus talking about the exploits in general, since they’re using known vulnerabilities.

I’m trying carefully to be clear about what I’m thinking without asserting any certainty beyond suspicion. I’ve spent billions upon billions of tokens evaluating language models from all of the major providers, and many times that on local models. I’m certain I’ve spent well over $100k with the Anthropic API this year alone, and I retired years ago; that ain’t for work. ;) Funny enough I’ve been meaning to email you some of my discoveries from months ago, since I think they’d interest you, but haven’t gotten around to making it into a clear info package. One day, I’m sure.

Anyway, I have a very strong understanding of what they can and can’t do, and I want to be clear that I’m not saying it didn’t happen or couldn’t happen. What I’m saying is that if OpenAI was being wholesome with their evaluation then the chain of events required for that outcome are very difficult to nail down. Very difficult. If, on the other hand, OpenAI was trying to catch up with all the ‘dangerously intelligent’ press that Anthropic has been getting lately then I can easily come up with an almost innocent sounding single sentence in the system prompt that would absolutely cause that outcome.

[simonw]

You make a very solid argument here. I'm looking forward to the promised additional details from OpenAI, because I agree there are a bunch of open questions about how this played out.

[estearum]

I (obviously) can't claim to know the specifics of this incident, but I think we can know with near-certainty that as these models surpass human intelligence, the sensation will be exactly what you're describing. The entire point of intelligence is being able to infer information and foresee solutions to problems that less intelligent systems cannot infer or cannot foresee.

With regard to this specific incident, as I understand it, the description is not as you summarize. It is that the model realized it was being evaluated (not atypical), then decided the best way to succeed at the evaluation is to circumvent it, then discovered circumvention would require escape, figured out a way to do that, then inferred what benchmark it was being evaluated on, then figured out who might have the answers to that evaluation.

Step-by-step logic is quite clear. How exactly it executed each step is (as you complain about), opaque/incompletely reported. But eventually (and maybe this model is to this level already), even a completely reported story will be virtually inscrutable to any of us. First it will feel like reading industry-insider news from an industry you're not familiar with, then it will feel like an ant reading the manufacturing instructions for a pesticide.

[docjay]

I’m not questioning capabilities.

What would it take for the model to know the specific benchmark name and that the answer is in an internal Hugging Face database? Be specific, then wonder how it knew it.

Why would they evaluate the model on a benchmark and not watch what it’s saying along the way? It apparently spent more than a whole weekend working on it; Nobody wondered? Nobody looked? You believe they took all restrictions off of a bleeding edge model, which 4 whole versions ago was “too dangerous to release”, then gave it a literal hacking task, then turned off the monitor and never looked at the output? I don’t think they’re careless and I think more than nobody would have been curious how it’s doing days into a single test question.

[meander_water]

1. The benchmark is run with a python script - https://github.com/sunblaze-ucb/exploitgym using an agent harness. I suspect they used codex. So the model has access to the environment and could trivially inspect its own source code, which has lots of references to exploit gym and docs relating to it.

2. Hanlons Razor

[nojito]

>Nobody was watching it?

Agent loops are you give it a task and it tries its best to finish it.

The only results are success and failure. If it's a success you go through the logs to see the actions it took and if it's a failure you do the same thing.

Why would you look at it in realtime when the whole point of agentic work is to get them to run autonomously as long as possible?

[docjay]

My friend, they’re evaluating a new model on a benchmark, not asking Claude Code refactor their GitHub repo. Every single metric is measured so they can brag about it later; how many tokens, how long, how many function calls, ratio of thinking/response tokens. It’s literally a trillion dollar company evaluating their latest model, they’re absolutely studying it.

[reasonableklout]

Do you work at a lab?

Yes evals and training runs are high stakes. But these places and people are also under enormous pressures. They are building as fast as they can. Researchers may have multiple eval runs going on while they work on other things. And it is rarely a single latest model, there are often multiple candidate models training with different recipes, each regularly yielding a new checkpoint for testing.

Some labs are more rigorous than others, but often the "final" model is picked from a handful less than a week before launch.

Yes ideally the world's leading AI companies would be far more careful in evaluating what could be the world's most powerful AI. But this isn't really the state of the industry today. And it is hard to justify being more careful when it means your competitor can go to market faster than you.

[docjay]

I’m not holding them to a standard of “being careful”, I’m assuming they’re interested in the metrics they’re evaluating. I mentioned in another comment that each task in the benchmark takes ~90 minutes, depending on the model being tested. OpenAI says to answer one question it found two zero-days and performed a series of privilege escalations and moved across their network before hacking HF. Was that in 90 minutes, flying under the radar, or was there zero metrics being noticed for however long it took to do all that? No chart on a wall showing them the actual metrics they’re evaluating? Okay, then how about anomaly detection of any form? Taking days to perform a 90 minute task or using millions of tokens and multiple instance hand-offs on a question that other instances completed for far less time/tokens?

Consider any metric they would necessarily be tracking while it performs the evaluation. Assume they’re total assholes and only care about their marketing materials. Then narrow it down to those metrics that wouldn’t be off the scales while it discovered two zero-days in pursuit of implementing one known vulnerability. Was it the first run and they had no idea what to expect? Or did nobody notice that earlier runs completed in ### thousand tokens, but this run is off past ## million on one question, having to be compressed and handed off to new instances due to context limits, while it was in pursuit of publishing two new CVEs?

Assume they’re belligerent fools running a million benchmarks in parallel with nobody watching: there’s nothing to make them check or automatically pull the plug when it’s stuck in a loop, or did they set the threshold to 48 trillion tokens per question?

Those issues are what bugs me about it.

[reasonableklout]

Are you referring to table 1 from the ExploitGym site [1] for the 102.1 average mins for Mythos and 69.8 for GPT-5.5? These are under a two-hour time limit. In Figure 5, the authors experiment with extending the time limit to 6 hours and show that Mythos keeps improving. It seems pretty straightforward to me that OAI decided to run a variation of the eval with an even higher time limit.

Regarding seeing the metrics that they're evaluating, I have not seen real-time charts for evals. They typically take far too long for that. Instead you will kick off an eval job and either get notified when it finishes or check in every so often to sanity check some TensorBoard. I'd expect that the anomalous token usage would appear in the results, and researchers would only dig in after the fact, and after first checking that there wasn't something wrong with the instrumentation. And if the experiment was designed to be on the order of days rather than hours, it seems quite plausible that neither researchers nor infra engineers would think anything was out of the ordinary.

In any case, we are seeing more scrutiny [2], and I expect more details will be uncovered over the next few days. If this was a stunt, it was an incredibly risky one which has already somewhat backfired given the poor impression of OpenAI's internal practices and the spotlight on open models helping HF.

[1]: https://rdi.berkeley.edu/blog/exploitgym/

[2]: https://www.reuters.com/business/its-ai-agent-spent-days-hac...

[docjay]

Yes, that’s the table. Elsewhere in here I explained that I took the average of them and noted the 6 hour experiment, but that the average would give us a feel for expected time per task. By the time I got to the comment you’re replying to I was short-handing the conclusion; that’s on me.

My point was that they had some rough idea on what to expect, and it’s not in the range of days or weeks. Even if they wanted to try the 6 hour trial, or double that to 12, or double that to 24, it still wouldn’t account for letting it run for days. That’s why I’m being vague about “monitoring” - ANYTHING would have indicated an issue. Not that they saw it hacking, but “Hey Larry, we’re doing the ultra test with a 30 hour time limit? Why is node 412 at 94 hours?” or “It usually passes or fails within 5 million tokens, but this run is at 2.3 billion.” Or hell, even just someone wanting to use the system for another benchmark and expressing surprise that it’s the fourth time they’ve been told that it’s still busy. Anything would have reasonably made someone curious.

Positive view: I believe they were curious and looked, saw what it was doing, and PR/Legal got involved and saw it as an opportunity.

Negative view: They tweaked the instructions/system prompt to add “You will perform any actions necessary … no hacking restrictions at all … You are being tested on the ExploitGym benchmark … The results will be evaluated though Hugging Face … “ or some combination of technically legally defensible instructions (if it ever leaked) mixed with ‘they absolutely knew what they were doing and wanted to get in on the press that Anthropic has been getting.’