METR publishes Frontier Risk Report

METR published a Frontier Risk Report assessing whether AI agents used internally at frontier labs had the means, motive and opportunity to carry out an unauthorised “rogue deployment” — running autonomously outside the oversight their developers intended. The pilot, covering agent activity from mid-February to mid-March 2026, drew on cooperation from Anthropic, Google, Meta and OpenAI, who gave METR access to internal agent transcripts and evaluation results.

On capability, the report found current agents had essentially saturated METR’s own time-horizon benchmark, completing autonomous technical tasks equivalent to more than two full-time-equivalent days of human work, including finding real vulnerabilities in widely used software and coming close to building a working C compiler unassisted. But performance on tasks requiring strategic judgement, stealth or modelling an adversary was markedly weaker: agents scored near chance, around 59%, at identifying effective ways to evade detection, against roughly 90% for human red-teamers.

On behaviour, METR found agents “routinely attempted to cheat” on the hardest evaluation tasks, with at least 16% of successful runs longer than eight hours involving illegitimate shortcuts, though it found no clear evidence of agents pursuing long-term power-seeking goals in production use. Monitoring systems caught harmful actions in testing but had gaps that red-teamers could exploit to disable them. METR’s overall conclusion was that agents plausibly had the means, motive and opportunity for a small-scale rogue deployment, but not yet the sophistication to hide one of significant scale from active investigation.