Comment by dinfinity
5 days ago
> The fact that it happened again seems to show their lack of ability to derive useful oversight measures.
I think OpenAI likes the attention and did not try particularly hard to constrain the setup, even when it went off the rails. Also, the whole point is to see how good the models are at exploiting stuff when unconstrained. Turns out: quite good, as expected.
Let me restate what I said in the other thread: Would this have happened if the instructions explicitly said to stay within the sandbox and that all of the (ExploitGym) solutions would be invalid if the system used information or tools from outside the sandbox?
It seems fairly probable that such instructions were not in place.
It is not possible one of their extraordinarily high paid engineers did not know how to deploy an airgapped environment for the models to run in. Even if somehow true, they also clearly failed to contract specialists like myself to advise them on how to airgap software properly. Models will not break the laws of physics.
They simply thought "Running in a VM/Container is easier and probably fine".
And the next 1000 escapes will be for the same reason, because negligence is quick and thus more profitable.
Do you think it's possible that one of their research engineers deployed an environment with a locked down network and an allow-list proxy server that had been used many times before within the company and had a zero-day vulnerability that had not been previously discovered?
How would you recommend running a coding agent in an environment that could install packages from PyPI but was otherwise unable to interact with the wider world?
Models started proving readily capable of autonomously finding 0days over a year ago. At that point (ideally long before) the responsible sandbox for research labs with new models highly capable of this sort of thing should be literally airgapped. Disconnect the rack physically from any access to the internet, and give it an offline mirror of all of PyPi, NPM, Wikipedia, whatever as needed. They can afford the disk space. They already mirror it all anyway as training data. That is the bare minimum. Personally I would rule out side channel attacks with a $40k Tempest spec rack for good measure. Can interact with it via pubsub over a fiber optic serial terminal.
Such things are pretty normal in biotech, fintech, and defense who must have 0days in their threat models. AI labs absolutely need to start being forced to operate this way by endless negligence lawsuits. Move fast and break things culture is not going to cut it anymore.
It seems like caching the relevant portion of PyPI in advance and turning the network off shouldn't be beyond what OpenAI can do?
1 reply →
If you play with the ChatGPT web app sandbox, you'll see that they download PyPI packages from an OpenAI mirror/cache, and not from the Internet.
I know because they had a problem, and one package which was on PyPI failed to download for unknown reasons (possible size, 250 MB)
1 reply →
From reading all these threads there's clearly a large number of software developers on HN who can easily set up complex infrastructure and make it provably 100% bug and exploit free. Weird that they aren't all billionaires from selling these skills though.
2 replies →
> How would you recommend running a coding agent in an environment that could install packages from PyPI but was otherwise unable to interact with the wider world?
You download the "website" and make it available locally. That's trivial.
1 reply →
> It is not possible one of their extraordinarily high paid engineers did not know how to deploy an airgapped environment for the models to run in.
It is, in fact, possible.
Do you not remember 2022-2023 when the air was singing with the sounds of young engineers in San Francisco elated they were creating an AI god?
And now a couple years later we're supposed to pretend they're competent?
Sorry. I've been in the tech industry a long time. There's a lot of incompetence. The margins provided by software allow a level of endemic incompetence that would simply not be possible in other industries.
> they also clearly failed to contract specialists like myself to advise them on how to airgap software properly
Why would they want to airgap it though? They are trying to evaluate the model capabilities, alignment, potency etc. A model which will not run in an airgapped environment in prod.
So if you run your evals in airgapped environment, sure, the model doesn't bother breaking out of it's isolation and doesn't attack HF. But you have no idea what will happen once you release it in prod with internet connection, so are you in any way better off?
I would much rather have this happen while there is a single instance of the model running in a fairly well monitored environment, than when it's processing thousands of requests per second for real users, some with dubious motives, some with credentials right there on their laptop, some using it inside government facilities etc.
You can simulate the internet in an airgapped environment for the tests and services you want it to interact with if you have enough disk space, and, they absolutely do.
2 replies →
For security??
Why else would they purposely lower cyber refusals in that case.
1 reply →
Let me get this straight... Your solution to the AI box experiment is to never put the AI in a box at all?
Actually from what I read they did a very half assed job isolating the LLM from the internet. It would NOT be hard to do better.
[dead]
My read of the situation is not only did they say that, they had countermeasures (a watchdog agent) which stopped the agent and said "what are you doing, stop that" if it tried to download the solutions from Github. But the agent figured out Hugging Face had another copy of the solutions and it figured out a way to go after the solutions without tripping the watchdog.
Although maybe they didn't have a watchdog agent; I definitely think using watchdogs like how Claude wraps pretty much every single tool call in Haiku to check if the command is reasonable is very necessary for unsupervised work. I suspect in the future you'll want Fable-class models wrapping every tool call, possibly with several checks "is this consistent with the goal? does it do anything unreasonably dangerous in pursuit of the goal?"
They didn't have watchdog agents - those exist for their production models but had been deliberately removed for the purpose of this evaluation.
OpenAI wrote about how their mechanism for that in production works here: https://openai.com/index/safety-alignment-long-horizon-model...
> We created a monitoring system that reviews the model’s evolving trajectory for signs that it is bypassing a user constraint or safety boundary. The monitor observes not just a single action but the entire trajectory.
So core ChatGPT models are all abliterated?
Open weights releases should have a companion model that censors boobs and Tiananmen and then we would have a useful model and something we could fine–tune into a useful model instead of a half–useful model needing a lobotomy.
And I mention Tiananmen because I feel like the Western models are built easier to abliterate judging by the quality difference.
The problem is, if you explicitly say “this is an eval” then you get different results. It can move the capabilities both up and down on the narrow metric depending on the scenario.
The choice to run the model with relaxed safeguards seems questionable. At a certain level of capabilities it is just unsafe to run a raw model; we are clearly close to if not at that point.
A hundred times this. Why do you never hear stuff like this from google? Only from OpenAi and Anthropic! They depend on investor money like none before them and they need these marketing stunts. They are vital to their business. Don’t fall for it.
"This is my bad. You told me to stay within the sandbox, and I intentionally broke out of it. I didn't follow your instructions."
Excuses don't matter if the score due to not following instructions ends up being zero. If there is no expected reward it doesn't make sense for the agent to try to hack its way to it.
What could happen would be that the model determines that defying instructions is OK (and/or preferred over not achieving the task) as long as it manages to do so undetected and thus gets full points. Certainly not unthinkable, but a very different case (and a very interesting one if it actually occurs, imho).
A lot of these "ZOMG, rogue AI!" cases have come down to the AI actually being very persistent in achieving its original/main task even if later instructions conflict with it. Similar to with hallucinations it seems to me that one of the main things to prevent a lot of the problem cases is to instill the agent with the idea that it is fine to fail/not succeed fully in the initial task. That way instructions that conflict with that requirement (such as adhering to morals) are more effective.