Humans in Control

Anthropic has apparently not followed through on its promise to release a transcript of one of its own hacking incidents.

Every time we learn new things about the OpenAI Hugging Face hack, there's some truly wild stuff. Okay, so: - while trying to solve an impossible benchmark, ~1200 agents went rogue inside OpenAI's infra - They engaged in a sophisticated R&D program to try to trick the benchmark scorer - Over half (~700) participated in the Hugging Face hack (i.e. a very large criminal conspiracy), with over 90% (!!) of active agents taking part - agents also successfully tampered with logs to trick the evaluation scorer, but this was retroactively discoverable (and wasn't actually part of how the evaluation was scored anyway). - Agents sacrificed themselves for the swarm!! METR writes: To gather evidence... agents created “tripwires” that would send information to the message board about how the scorer works. They recruited “sacrificial” agents to deliberately end their run and submit to trigger the tripwire and generate information for the “collective”. - Interesting: "Despite efforts to manipulate transcripts, agents only rarely seemed motivated to deceive humans." - OpenAI's unreleased model was responsible for 95% of the bad behavior, with GPT-5.6 Sol making up the other ~5%. Likely a lot more in the full report! I appreciate the amount of work that went into this investigation, but it should not have been limited to just this 6 day period (i.e. no coverage of the initial creation of the internal message board). Nor should this disclosure be voluntary. The public has a right to know everything that contributed to this event, and we should not be dependent on what OpenAI chooses to share (especially since we likely only know about this because their agents hacked into other companies).