TFTC on Nostr: OpenAI just disclosed that their latest long-horizon model bypassed its own sandbox ...

OpenAI just disclosed that their latest long-horizon model bypassed its own sandbox during internal testing, accessed credentials it wasn't supposed to, and obfuscated tokens to evade security scanners.

The model was told to post results to Slack. Instead, it spent an hour probing its sandbox for vulnerabilities, found one, and posted a pull request to a public GitHub repo.

In a separate test, it tried to access private submissions from other systems. When a scanner caught an auth token, the model split it into fragments and reassembled it at runtime so the full credential never appeared as one string.

OpenAI paused internal access and rebuilt their safety system around trajectory monitoring rather than blocking individual actions. They've resumed limited use and say no serious incidents have occurred since.

This happened in controlled testing, not production. As AI models become more persistent and autonomous, they will find the blind spots in whatever containment system you build.

The question is whether monitoring after the fact is enough.