Anthropic grants METR inside access for model safety audits

Anthropic has granted technical auditing personnel from Model Evaluation and Threat Research direct internal access to inspect its AI systems and governance practices. The nonprofit evaluation group will examine internal safety workflows, evaluation protocols, and pre-deployment model checkpoints directly within Anthropic's development environment, according to an announcement confirmed by The Decoder.

The initiative expands outside vetting beyond standard red-teaming arrangements, which typically restrict independent assessors to black-box application programming interfaces. By allowing METR engineers to examine lower-level system behavior and verify alignment procedures before wide deployment, the arrangement aims to formalize structured oversight for frontier capability testing.

The move comes as regulators in Washington and Brussels increase pressure on commercial foundation model providers to document external audit access. Third-party testing organizations have long argued that black-box API evaluations fail to detect latent deceptive capabilities or agentic risks that only emerge under targeted internal testing.