Anthropic has granted nonprofit evaluator METR employee-like access to its hardware, internal systems, and offices to assess safety risks in its AI models. According to the company, six METR staff members have been onboarded with hardware and access privileges on par with internal risk assessment teams. The arrangement is uncompensated to preserve independence, and METR retains the right to publish its findings without an Anthropic veto, subject to redacting proprietary architecture and customer data.

The move fulfills commitments made alongside industry leaders to establish third-party auditing for frontier systems as Anthropic prepares for a potential initial public offering this year at a valuation exceeding $2 trillion. Critics, however, argue that voluntary auditor relationships present potential conflicts of interest and that formal government regulation remains necessary. Meanwhile, Anthropic continues safety testing partnerships, including a joint $1 billion commitment with Accenture.

Why it matters

  • Sets a benchmark for independent frontier model auditing ahead of Anthropic’s expected IPO and heightened regulatory scrutiny.

  • Demonstrates how leading labs are operationalizing external oversight through embedded, uncompensated third-party evaluation teams.

Source: arstechnica.com