Frontier Risk Report (February to March 2026)

Testing a model before public release can miss risks from how its developer already uses it internally. METR's February–March 2026 assessment combined model access with information about internal use and safeguards at Anthropic, Google, Meta and OpenAI. It examined whether agents had the capabilities, incentives and opportunity to operate without human permission.

METR judged that small unauthorized deployments were plausible during that assessment window, while highly resilient ones exceeded the assessed agents' capabilities. This was a risk judgment, not a report that such deployments had occurred. Participants could redact their non-public disclosures or withdraw before approving them; they did not have approval rights over the final report. Those terms matter when weighing the access gained and the limits of outside scrutiny.