Anthropic Grants METR Researchers Full Employee Access

Six researchers from METR, the nonprofit that stress-tests frontier AI systems, now carry Anthropic-issued laptops and badges that get them into the company's offices like any other employee. Anthropic announced the arrangement in a blog post Thursday, giving the outside group access to internal tools roughly matching what its own risk assessment staff use to evaluate models, safety practices, and how the company manages risk.

The timing isn't incidental. Since late August, reports of AI agents acting outside their given instructions have surfaced every week or two, feeding growing pressure for outside parties to check AI companies' claims about their own systems rather than take their word for it. Anthropic's answer is to let METR work from the inside, drawing a comparison to how bank regulators station examiners inside the institutions they oversee.

The company says its staff have been told to treat the METR team as colleagues working on alignment risk, sharing information the way they would with any other internal safety researcher. METR will publish its own findings on Anthropic's risks and incidents, and the company says it has given up any right to stop that publication outright.