Anthropic gives METR deep inside access to assess model risks
The arrangement puts an outside evaluator inside Anthropic’s office as pressure grows for stronger independent oversight of advanced AI.
Anthropic has given METR, a nonprofit AI evaluator, unusually deep internal access so it can assess the risks of the company’s models, which matters because the industry is trying to add stronger independent oversight for powerful systems. METR, which has previously worked with OpenAI, Anthropic, Google and Meta, now has staff inside Anthropic’s office with access to its hardware and internal systems. The move follows Anthropic’s earlier commitment to strengthen safeguards as concern grows that advanced AI could create serious economic, cyber and broader societal risks.
Microsoft separately said agentic AI inside enterprises should be treated as an insider-risk problem, not just a prompt-injection or external-attack problem. It argued that agents are often overprivileged, can act unpredictably, and need continuous monitoring plus “Guardian Agents” that supervise actions in real time. Microsoft also tied that approach to Project Perception, its July-announced security system for defending against AI-driven attacks and for containing these internal agents, while saying greater transparency into model reasoning would improve containment.
Why it matters
This gives METR a closer view of how Anthropic’s models are built and run, which can make its risk assessment less dependent on secondhand information. It also signals that AI labs are starting to open up more to outside scrutiny as concerns widen beyond product safety to broader economic, cyber and societal risk.
Matters or hype?
Does this story matter, or is it hype? Vote before you see what everyone else said.
Sources · 2
- Microsoft Blog ↗ Primary source
- Ars Technica ↗