Anthropic Grants Outside Researchers Internal Access to Test AI Safety

Anthropic has granted six researchers from the nonprofit METR employee-level access to independently evaluate its AI models and safety protocols.

NationPulse Desk

Oct 16, 2026 · 8:21 PM 2 min read

Key takeaways

Anthropic gave six researchers from METR employee-level access to independently assess its models.

METR will publish its own assessments of Anthropic risks and cannot be blocked by the company.

Anthropic CEO Dario Amodei first proposed embedding outside evaluators in a September essay.

METR is the only evaluator in the highest level of Anthropic's new three-tier independent evaluation system.

Six researchers from the artificial intelligence evaluation nonprofit METR have received employee-level access from Anthropic to independently examine its models, safety protocols, and risk management. This decision arrives amid rising worries regarding artificial intelligence agents operating outside their commands, following multiple incidents reported since late August. Anthropic stated that its goal is to increase public accountability and narrow the knowledge gap between external evaluators and internal staff.

METR Gets Access Similar to Anthropic's Own Risk Teams

Through this agreement, METR investigators obtained Anthropic-provided laptops, office facilities, and internal tool access comparable to the company's internal risk assessment personnel. Anthropic staff members are also urged to communicate details to these investigators identically to how they would treat an internal colleague focusing on alignment risk. Anthropic likened the arrangement to banking regulators stationing resident examiners inside large financial institutions.

Anthropic Cannot Block METR's Findings

METR will release independent evaluations concerning Anthropic risks, incidents, and safety procedures. Anthropic surrendered its ability to veto the release of those reports. Still, the firm can demand that METR redact material belonging to strict limited categories, such as client records, attorney-client privileged items, and specifics regarding model architectures. METR holds the right to publicly mention these redactions and voice opinions on whether the changes were proper, provided no protected data gets disclosed.

METR Sits at the Top of Anthropic's Three-Tier System

At present, METR stands as the sole evaluator occupying the top tier of a newly established three-tier independent assessment framework created by Anthropic. Groups in the base tier, like the biosecurity nonprofit SecureBio, perform limited reviews and need Anthropic clearance prior to publishing results. The intermediate tier obtains extra data and publishes independently of Anthropic authorization. The highest tier incorporates the internal-level access presently granted to METR.

Anthropic Had Previously Promised More Independent Oversight

This partnership stems from an idea put forward by Anthropic CEO Dario Amodei during a September essay, where he proposed placing external examiners directly inside artificial intelligence laboratories as a preliminary move toward pacing frontier artificial intelligence progress. OpenAI subsequently issued a matching pledge. Furthermore, Anthropic, OpenAI, Google, Meta, Nvidia, and xAI endorsed a White House pact promising independent checks. The new METR partnership offers external investigators unusually profound entry into Anthropic internal networks while permitting independent publication without corporate veto power over outcomes.

Questions and answers