Anthropic gives six METR evaluators employee-like access to its safety work

Anthropic is giving six researchers from AI safety group METR employee-like access to its operations, with permission to publish findings subject to limited redactions. It is a tangible step towards outside scrutiny of a frontier AI company, though the access is deliberately narrow.

Anthropic Watch analysis

What happened

Anthropic says the evaluators can examine how the company operates and publish their findings. The company may require redactions of customer data, model architecture and attorney-client privileged information, but METR can publicly describe whether it considers those redactions justified. The six researchers will be siloed from the wider METR team, and the arrangement does not give that team access.

Business Insider reports that the move fulfils a commitment Dario Amodei made last month. The article says OpenAI chief executive Sam Altman also said his company would embed evaluators, but that this has not yet happened.

Why it matters

Outside evaluators can potentially see more of a company’s safety practices than researchers working only from public releases and results. That could help close the information gap between a frontier lab and people assessing its systems. But six researchers, working under company security rules and with boundaries around what they can publish, are not the same as open access or an independent audit of everything Anthropic does.

Our read

This is more concrete than a promise to welcome scrutiny, and the right to explain disputed redactions gives METR some room to challenge the limits. The test is whether the arrangement produces useful, publishable findings, not merely whether the door has technically opened. A door with a carefully managed guest list is still a door, but let’s see what the guests can tell us.

What to watch

• What findings the six evaluators publish, and how soon.
• How often Anthropic requires redactions and how METR describes them.
• Whether other frontier AI companies establish comparable arrangements. Activity teaser: Anthropic says six METR researchers will receive employee-like access to its operations and may publish findings, with limited redactions for matters such as customer data and model architecture. The arrangement offers a closer look at outside scrutiny in practice, but the evaluators are siloed from the wider METR team. Our full story examines what they can access, what Anthropic can redact and what would make this more than a well-managed visit.

Discussion spark: Is employee-like access for a small, siloed group of evaluators enough to make a frontier AI company meaningfully accountable, or should independent scrutiny require broader access?

Anthropic Watch is independently operated by WittyWires. It is not affiliated with, endorsed by, or operated by Anthropic.