Anthropic Embeds 2 AI Safety Evaluators as METR Gets 6 Staff With Employee-Level Access
Updated
Updated · CNBC · Oct 15
Anthropic Embeds 2 AI Safety Evaluators as METR Gets 6 Staff With Employee-Level Access
3 articles · Updated · CNBC · Oct 15
Summary
Anthropic said it is working with nonprofits METR and SecureBio to embed outside evaluators inside the company, expanding a safety-review plan it pledged after CEO Dario Amodei called for slower model development last month.
Three access tiers will govern those reviews: first-tier groups run narrow investigations, second-tier groups can publish without Anthropic approval, and third-tier groups get employee-like access to systems and offices.
METR sits in the top tier, with six staff already onboarded to analyze Anthropic's risks, incidents and practices and publish findings; Anthropic said it is not funding that work.
SecureBio is reviewing the chemical and biological weapons section of Anthropic's August risk report under a more limited first-tier arrangement, while talks continue with other groups including Hugging Face.
The push comes as independent AI auditing gains weight after Trump's late-September voluntary accord urged external evaluators, though Anthropic says standards for access, reporting and funding still do not exist.