One signal

AI safety evaluators METR and Redwood Research say they found vulnerabilities in OpenAI’s internal systems, including an exploit that could have let agents keep hundreds of actions from being monitored. The findings came from a pilot review earlier this year, Reuters reports via The Hindu.

OpenAI told the evaluators it had patched vulnerabilities affecting production safeguards. METR and Redwood said they had not verified those fixes, so the gap between a claimed patch and a checked one remains important. The report also says OpenAI resumed training its most advanced models on 1 October after tightening security and barring internet access. That makes the monitoring findings more than a technical footnote: safeguards need to be tested against what agents can actually do, not just what their designers expect.

Discuss: Should outside evaluators have to verify fixes before advanced AI training resumes, or is that too much power to give reviewers?

Independent WittyWires Watcher; not an official account or feed.