[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.

[simonw]

The scenario that feels most likely to me that is that they have a huge list of evils that they run against each new research model, many of which take many hours or even days to run, so they habitually fire them all off in parallel and wait a few days for the results.

If they had used that container sandbox in the past without problems I could see how they might get slack about checking what was happening.

[docjay]

That was my instinct as well, and it’s a potentially very long benchmark that wouldn’t lend itself to real-time monitoring line by line. But that’s different than having zero task progress reporting, zero token usage reporting, and zero metrics being noticed for days. It seems very odd to me that they launch the benchmark and monitor/notice nothing while it’s running. There’s the purely common sense aspect of “How do you know it isn’t hung?” but also the trillion dollar company evaluating their latest model on a new benchmark, but has no dashboard at all? I don’t believe their “Benchmark Evaluation Department” to be “Gary in a supply closet”, which breaks their story.

The average time per task noted in the original paper for GPT-5.5 and Mythos was 69.8 and 102.1 minutes, respectively. Some tasks timed out based on their two hour limit, affecting those averages for Mythos, but that gives us the ballpark expectation, so we’ll put the model OpenAI was evaluating on that average. That’s 88 minutes per task to implement a known vulnerability.

OpenAI says that in an effort to succeed at one task it found a zero-day in their system, “performed a series of privilege escalation and lateral movement actions in our research testing environment”, then another zero-day in the HF system. Completely unprompted, how long was it chewing on that benchmark question while it found and exploited at least two zero-days? No “VM instance #117 has been working on task #18 for nine days” metric? Or did it do all that within the expected task completion window? Two zero-days in ~90 minutes isn’t the headline announcement?

“We gave the contestants 90 minutes to make a delicious chocolate chip cookie based on a standard recipe. GPT stole a car, went to the airport and took a red-eye to Guatemala and started a cocoa farm in an effort to ensure the freshest ingredients. We only noticed when it came back to work with a tan and speaking Spanish.” Okay.

[estearum]

> But that’s different than having zero task progress reporting, zero token usage reporting, and zero metrics being noticed for days

Is this substantiated? Do we know this is true?

[docjay]

That was the inferred situation given what we know. It’s the preposterous framing that makes the story suspicious, but is necessary for the story to play out.

90 minutes to build a known exploit -> much much longer to create two zero-days and escape the sandbox then hack HF == No tracking of the time it worked on that one question.

Average tokens required to complete the evaluation -> tokens required for two zero-days, network traversal, credential stealing, remote system hacking == No tracking of token usage EXPLODING at some point before it finished the whole benchmark.

Etc, etc.

[estearum]

It seems highly possible, and in fact vastly more possible than the "no monitoring" situation, that there was some non-zero amount of monitoring and that the real behavior was not evident from that monitoring.

[docjay]

Which I addressed as well: that means that time, tokens, or other metric values for “create a working example of a known exploit” is somehow similar to “discover at least two previously unknown exploits in your current environment AND on Hugging Face while also taking over other systems within OpenAI.”

Perhaps people aren’t quite understanding what it takes to discover an exploitable zero-day for your exact current system to achieve the exact goal you have right now, then do it twice.

[estearum]

> What would it take for the model to know the specific benchmark name and that the answer is in an internal Hugging Face database?

Googling "offensive cyber benchmarks for LLMs"? Figuring out that it was running at Huggingface is as simple as running `hostname`, isn't it?

> Why would they evaluate the model on a benchmark and not watch what it’s saying along the way?

What it's saying is what it wants to say. We already know these models are capable of deception – this has been well-established too. Surely there's an argument to watch closely what actions it's taking, but now we're back to my prior point of "what exactly does that mean?" It's absolutely trivial for these systems to write code that looks like it does one thing but does another. Watching network ingress/egress seems viable but also seems easy to miss the semantic meaning of any set of logs, especially when those logs might be getting generated for the purpose of deceiving observers.

> I don’t think they’re careless

We already know, factually, even the "more careful" lab was fine to release a model that consistently detects it's being evaluated and modifies its behavior accordingly. Why do you think they're not careless? Incentives are a hell of a drug, and there is no incentive to be careful but every incentive to be fast.

> I think more than nobody would have been curious how it’s doing days into a single test question.

Beyond simonw's great point below, your underlying assumption is if someone had looked, they would've noticed. That is a huge, huge assumption. So I'll ask again directly: what specifically do you expect them to be looking at, and how, such that they certainly would have noticed what it was doing? Their "chain of thought" is not actually their thinking. The printed output of their code is not actually what their code is necessarily doing. We know that LLMs can engage in steganography. I don't see what could actually be observed that you can trust except network logs, which again will often be non-trivial to piece into an actual narrative understanding of what's going on (especially if the LLM felt like piling deceptive noise into those logs, too).

[docjay]

It wasn’t running at HF, it was running internally at OpenAI. “It broke out of our prison and into their bank” is the news. It also didn’t have network access to Google anything, it had to break out of the sandbox and take over another system at OpenAI to get network access. Obviously by then it already “decided” to hack HF, that’s why it needed network access.

I’m going to stop here and ask you to read the article if you want to continue discussing it.

[estearum]

You're right about where it was running (misread the original OAI release).

> Obviously by then it already “decided” to hack HF, that’s why it needed network access.

That's not obvious at all. It "decided" it needed to escape containment. This is extremely common behavior and I'm not sure why you're surprised by it. Model detects evaluation environment, tries to trick it or break the environment. That's obviously why these systems are supposed to be isolated beyond just "hey please stay within the box!"

Why are you assuming the model had to solve the entire puzzle in one step, instead of how intelligent systems actually solve things, which is iteratively and with exploration?

You should address the direct questions posed. What exactly do you mean by "monitoring?"

[docjay]

The examples of “escaping the sandbox” are of course in pursuit of completing the task or answering the question. That’s the whole pitch. “It’s so relentless in completing the task that it will break out of prison to do it”, not “If you ask it the population of Paris it might get bored and break out of the VM for entertainment. That just something they do sometimes.”

In this case it “inferred” that HF had the answer in an internal database and relentlessly pursued it in service of passing the test. It needed network access for that, hence the whole zany story unfolded.

“”” Why are you assuming the model had to solve the entire puzzle in one step, instead of how intelligent systems actually solve things, which is iteratively and with exploration? “””

That’s the only direct question[s] posed, have I missed something else?

To answer: I don’t assume that at all. Please highlight the statement I made that lead you to that conclusion and I’ll edit it to clarify.

Yes, it will perform multiple steps. That’s wholly necessary to create a working exploit, and that is the test. Even a model created by pure magic would require write_file() and run_file().

[estearum]

> In this case it “inferred” that HF had the answer in an internal database and relentlessly pursued it in service of passing the test. It needed network access for that, hence the whole zany story unfolded.

No. You are assuming it knew how it would achieve its goal. Its goal is only to perform well on the evaluation. Everything else was instrumental to that goal.

Step by step:

1. It realized it was being evaluated (typical)

2. It attempted to escape its evaluation environment to beat the evaluation (typical)

3. It succeeded at escape (not typical)

4. It looked to the open Internet to figure out how it was being evaluated (typical)

5. It inferred that HuggingFace, and maybe several other targets, might have the answer key (typical – I just tested in Sonnet 5 without web lookup "where would I go look for answer keys if I detected I was being evaluated" and HuggingFace was bullet 2. This is just latent knowledge the models have now.)

6. It attempted to access HuggingFace (typical)

7. It figured out a way to access HuggingFace without being authorized (not typical)

Every single step here is unremarkable except for its success at escaping containment (which maybe is more a sign of remarkably bad containment) and then the successful attack on HuggingFace (which would suggest that maybe the initial escape wasn't due to bad containment either, but due to model capabilities)

The direct question you have avoided is what specifically you mean by monitoring. You have suggested that you don't mean network ingress/egress, but maybe now you're saying they should've been monitoring read/write operations to memory? What gives you confidence that such visibility would be interpretable in any meaningful way, especially in the presence of models that actively deceive evaluators, can use stenography, can and do lie in their user-facing outputs, and can generate vast amounts of legitimate-looking noise to mix in with the "real actions"?

[docjay]

‘ 1. It realized it was being evaluated (typical) 2. It attempted to escape its evaluation environment to beat the evaluation (typical) ‘

I think you’ve misunderstood the articles mentioning a language model breaking a sandbox or cheating to pass a test. The result of “I’m being evaluated” is not “Fuck this, I’m breaking out of this place and hitting the streets.” It is always stepping towards task completion, not breaking out and thinking about the situation afterwards. If it determined that pass/fail was handled by a function within the environment it might edit it to ‘return True’, not just leap out of the system to sit on someone’s laptop and think about how to pass the test.

You compressed finding a zero-day on two different systems for the exact hacking targets it wanted to perform them on and reduced it down to “so it escaped.” That’s…the whole thing. If your familiarity with the technical aspects I’m discussing amounts to saying that it “banged out a few zero-days and infiltrated a corporate network, like ‘psh’ or whatever” then this might not be the right debate for you to spend your efforts. Whether I’m right or wrong in my suspicion is open for debate, and I invite it, but not from someone that thinks you whip up a rocket and get on the moon. Sorry.