Most companies treat outside auditors like in-laws at Thanksgiving: polite, tolerated, and kept far away from the good silverware. Anthropic just gave its AI safety evaluators a desk, the internal tools, and permission to publish what they find.

Six Desks, Three Tiers, Zero Vetoes

Anthropic announced it is embedding nonprofit evaluators inside its operations, starting with METR and SecureBio, and is in talks with Hugging Face, Tekedia reports. METR gets the broadest, top-tier access. Six of its staff are onboarded with Anthropic's hardware, office space and tools on terms "roughly on par" with Anthropic's own risk-assessment teams. METR plans to examine risk levels, incidents and practices, then publish what it finds. Anthropic says it isn't funding METR's work.

SecureBio starts at the first, more limited tier. It's reviewing the chemical and biological weapons section of the risk report Anthropic released in August. Accenture is listed as Anthropic's first partner, and Anthropic will fund its work directly, though the company hasn't said how Accenture fits into the three-tier framework. "We're embedding evaluators voluntarily, and we believe every frontier AI company should do the same," Anthropic said.

Trust, but Let Someone Else Verify

This is the AI industry slowly moving from "trust us, bro" to "check our work." Independent verification is exactly what's missing when every lab grades its own homework. It also comes after a voluntary accord in late September in which AI companies were urged to use independent external evaluators, which Anthropic signed.

The part people miss: Anthropic has itself said the pooled or government funding for evaluators it called for in June doesn't exist yet. That makes the "who pays the auditors" question the real test of whether this model spreads across the industry or stays a one-company experiment. Still, inviting watchdogs inside, with the right to publish, sets a bar competitors will be asked about.

In AI safety, the most reassuring words are now "we let someone else look."

Bringing AI into your business and wondering how you'd vet it with the same rigor? We help teams add AI with real guardrails and oversight, not just a pinky promise.

Source: Tekedia