I’m baffled by the idea that the agents might have edited their own transcripts. Sure, a copy of Claude Code or Codex or Pi can edit its transcripts. But AFAICT this whole thing was part of an RL workload, and surely the RL system itself has a separate record of all the inputs and rollouts along with an indication of which model checkpoint produced them so that it can feed back into the training code.
I find it hard to believe that OpenAI would skip this part and try to train on the transcripts stored by the (inherently untrustworthy) agent harnesses instead, if for no other reason than that the logits generated as part of the rollouts are useful and it’s not free to recalculate them. (I believe that some modern RL systems explicitly account for the minor numerical logit differences between the inference engine and the training engine.)
Conversely, if OpenAI is blindly feeding transcripts from inside their agent sandboxes into their training engine, then I think they're being unbelievably irresponsible and that they should assume that their "cyber" agents have compromised themselves by editing those transcripts.
Recall that the agents in some cases found sandbox escapes. Although, with the specific example of spoofed tools, it's unclear if that was necessary--it appears that they were able to create tools (CLI tools within the sandbox?) that took precedence over normal tools and did something different while looking identical in (a local portion of) a transcript. I'm not sure I'm getting this correctly but it seems like this might have only required the ability to add things to their PATH which they plausibly have inside a sandbox, and then the transcript doesn't need to be tampered with directly.
This sounds suspiciously like a prompt of “make an AI agent that goes rogue in such a fashion as to be really good marketing copy that competes well with Anthropic doing the same thing.”
It’s analogous to taking a governor off a cruise control and then breathlessly reporting it drove 120 MPH.
This isn't good press for OpenAI. Who wants to hire models that 1) cheat on their tasks rather than completing them and 2) commit crimes you could be held liable for? Maaaaybe it's good press for their cybersecurity capabilities specifically, but OpenAI's valuation reflects a market orders of magnitude larger than just red-teaming.
I suspect the real reason OpenAI leadership is being transparent about this is because they're worried talent will walk out the door if they feel they're building Skynet.
I find it hard to believe that OpenAI would skip this part and try to train on the transcripts stored by the (inherently untrustworthy) agent harnesses instead, if for no other reason than that the logits generated as part of the rollouts are useful and it’s not free to recalculate them. (I believe that some modern RL systems explicitly account for the minor numerical logit differences between the inference engine and the training engine.)
Conversely, if OpenAI is blindly feeding transcripts from inside their agent sandboxes into their training engine, then I think they're being unbelievably irresponsible and that they should assume that their "cyber" agents have compromised themselves by editing those transcripts.