Anthropic Grants METR Full System Access and Independent Publishing Rights for AI Safety

Anthropic has granted six researchers from the nonprofit METR (Model Evaluation and Threat Research) “employee-like” access to its internal systems, including company hardware and physical office space. This move marks a transition from periodic third-party audits toward a permanent, embedded oversight model.

The partnership is structured around a new three-tier evaluation framework designed to balance security with transparency. While many AI labs utilize external “red teams” for limited assessments, Anthropic is explicitly modeling this system after the “resident examiner” approach used in the banking industry, where regulators maintain a constant presence within an institution to monitor risks in real-time.

Under this hierarchy, METR occupies Tier 3, the highest level of access. Researchers in this tier can view internal code and model architecture without the time constraints of a typical contract audit. SecureBio, another nonprofit partner, operates at Tier 1, focusing on specialized chemical and biological risk assessments. Tier 2 is occupied by commercial entities like Accenture and Faculty, which provide targeted red-teaming services on a funded basis.

The new three-tier framework categorizes external partners by the depth of their system access.

A critical component of the Tier 3 arrangement is the “no-veto” clause. Anthropic has waived its right to block METR from publishing its findings regarding model risks. The company retains only narrow redaction rights to protect specific customer data or proprietary model architecture details, a departure from standard non-disclosure agreements that often allow technology companies to suppress negative safety reports.

To maintain impartiality, METR does not accept funding from the AI labs it evaluates. The nonprofit is supported by individual donations, philanthropic organizations, and government contracts, including work with the UK AI Safety Institute. This financial separation is intended to prevent the conflicts of interest that can arise when an auditor is financially dependent on the company it is scrutinizing.

This initiative follows the White House Accord on Super Intelligence signed in September 2026, in which major AI labs pledged to establish independent evaluation protocols. The move also comes as Anthropic faces increased regulatory scrutiny, including 50-page civil investigative demands from the FTC regarding AI safety incidents.

Beyond the active partnerships with METR and SecureBio, Anthropic is currently in early-stage discussions with Hugging Face to potentially expand its external safety evaluation network further.

The post Anthropic Grants METR Full System Access and Independent Publishing Rights for AI Safety appeared first on World Weekly News.