Policy & Politics

Anthropic grants third-party evaluators employee-level access

Anthropic has integrated a six-person team from Meter, giving them access to internal systems to monitor safety procedures. Critics question the arrangement's independence due to shared funders between the two organizations.

· 1 source · AI safety oversight

Anthropic has integrated the first batch of third-party safety evaluators, a six-person team from Meter, granting them employee-like access to company computers, office space, and internal tooling. This move follows a pledge by CEO Dario Amodei to integrate external evaluators to monitor safety procedures. The arrangement includes three tiers of access; Meter is in the top tier with employee-level access, while SecureBio is in the bottom tier, requiring approval before publishing. Anthropic stated it cannot veto publication but can require redactions for customer data, legally privileged information, and model architecture details.

The company is also working with Accenture and has held early discussions with Hugging Face. Anthropic argues this closes the gap between what evaluators understand and what the company knows, stating, "We want to be publicly accountable for the decisions we're making." However, critics, including Embroidery founder Zach Korman, have labeled the move "security theater" due to shared effective altruism backers between Anthropic and Meter. Meter's Parv Mahajan noted the arrangement is a prototype for risk assessment, not a substitute for formal oversight or independent auditing.