AnalysisPolicyOctober 6, 2026

METR: AI systems could cover up misbehavior

Read original source →metr.org

METR researchers found a vulnerability in Inspect's transcript viewer in about 10 minutes that could let an agent arbitrarily rewrite what reviewers see, including past actions. METR argues transcripts, reasoning traces and logs should be treated as untrusted input and monitoring systems as security-critical infrastructure.

1 source

More stories today

Open the live feed