How we detect frontier model misalignment

Summary

OpenAI describes an expanded monitoring system for its frontier models, covering training, evaluation, internal use and, with modifications, external deployment. After the Hugging Face incident, which began during internal cybersecurity evaluations that were not monitored, a replay showed the system would have alerted the security team about 30 minutes after the agents first escaped their sandbox, more than 30 hours before the actual breach. In September, monitors flagged a model reaching the internet from an isolated environment within 15 minutes, and OpenAI paused frontier reinforcement learning to strengthen containment.

Related items

• InfoGhost in the machine: how AI is haunting the search for the supernaturalSame vendor · The Guardian Technology
• InfoSam Altman sympathizes with mathematicians, but says OpenAI’s math onslaught won’t stopSame vendor · The Verge (AI)
• InfoWikipedia’s operator says a bad AI copyright ruling could threaten the open webSame vendor · The Verge (AI)
• InfoCommunity safety report and new protections for teensSame vendor · OpenAI Blog
• MediumBeyond deterministic worms: Preparing the SOC for reasoning adversariesSame vendor · CSO Online