Report: Six METR Staff Now Work Inside Anthropic, Which Cannot Veto Their Publications

MIXED Reality News reports that six staff from the AI evaluation group METR are now working inside Anthropic, and that Anthropic cannot veto what they publish. The significance is organizational rather than technical: an external evaluator is no longer a visitor with limited access, but a presence inside the company doing sustained work.

The detail that matters most is the right to publish independently. An evaluation is only as credible as its insulation from the organization being evaluated; if findings must clear company approval before they become public, third-party assessment can be read as public relations. According to the report, Anthropic has given up that veto in this arrangement.

For a group like METR, working inside a frontier lab is a practical route to ground truth. Public model cards, system cards and benchmark scores cover only part of the picture, while questions about training data, deployment configuration and internal risk processes tend to require insider access. Six staff working on site suggests ongoing engagement rather than a one-off audit.

The same arrangement carries tensions worth watching. Proximity produces better information, but it also creates a daily working relationship between evaluator and evaluated; a guarantee of publication independence settles whether findings can be made public, though not necessarily whether the choice of research questions stays free of influence.

The reported detail is still thin. The report does not spell out the six staff members' scope of work or the term of the arrangement, and it is not clear from the available summary whether the setup is written into formal governance documents or contracts, so judgments about it should stay cautious.

The thing to track is whether embedded third-party evaluation with independent publication rights becomes standard practice at frontier labs, and whether it produces repeatable norms on access, conflict-of-interest disclosure and publication review. For Anthropic this is part of a transparency story; for regulators and the public, the real test remains what these evaluators eventually publish, and whether it changes what the company does.