Anthropic Opens Its Doors: Inside the New METR Audit Deal

According to TechCrunch, AI safety isn't just about internal testing anymore. Anthropic has placed six researchers from the nonprofit METR directly inside its operations with company laptops, office space, and employee-level access to internal tools.

How the Three-Tier Evaluation System Works

TechCrunch reports that METR sits at the top of a new three-tier evaluation structure built by Anthropic. Lower tiers involve narrow reviews requiring company approval to publish, while METR's top tier grants deep system visibility and the freedom to publish independent risk assessments without prior veto power.

While Anthropic can still request redactions for narrow categories like customer data, legal privilege, or model architecture secrets, METR retains the right to publicly question whether those redactions were justified.

What This Means for Enterprise AI Adoption

For organizations relying on frontier models, independent oversight directly impacts risk management strategies. Knowing that external watchdogs have unhindered visibility into model behavior helps compliance teams justify deployment decisions.

Verdict: Who Benefits From This Transparency?

This move is vital for enterprise leaders and safety officers who require verifiable risk assessments rather than corporate self-policing. While it does not eliminate unexpected agent behaviors overnight, it forces a higher degree of accountability across the industry.