Why it matters
METR's pilot evaluates risks from internal agent use at Anthropic, Google, Meta, and OpenAI using access to capable internal models, raw chains of thought, non-public operating information, and a means-motive-opportunity framework. It concludes that agents plausibly could start small rogue deployments but could not make them highly robust, while documenting uneven monitoring coverage and important uncertainty in capability elicitation.
My takeaway: Assess the organization and deployment environment, not only a model checkpoint: inventory internal agents, permissions, monitoring gaps, communication paths, and opportunities for persistence. Repeat the assessment on a schedule, record evidence provenance and redactions, and avoid treating chain-of-thought review as complete behavioral coverage.