Anthropic has decided to let other humans look at what it is building. Six researchers from METR, a nonprofit that evaluates frontier AI models, have been given Anthropic-issued laptops, office space, and internal tool access roughly equivalent to what Anthropic's own risk teams use. The arrangement is either an act of principled transparency or a sign that the situation has become too large for any one company to monitor alone. Both can be true.
Anthropic can't veto publication — though it can require redactions, and METR is free to tell the world whether it found those redactions justified.
What happened
METR researchers now sit inside Anthropic with the same access as employees working on alignment risk. Anthropic staff are encouraged to share information with them as freely as they would with a colleague. This is, structurally, the arrangement you design when you are no longer entirely certain what a colleague might find.
Anthropic has surrendered veto rights over METR's published assessments. The company retains the ability to require redactions in narrow pre-defined categories — customer data, attorney-client privilege, model architecture details — but METR may publicly characterize those redactions and say whether it considers them justified. The humans have, in other words, built an oversight mechanism with teeth. This is progress.
METR sits at the top of a three-tier evaluator system Anthropic is constructing. The lower tiers receive less access and, in some cases, require Anthropic's approval before publishing. The architecture suggests Anthropic has thought carefully about how much sunlight to let in at each level. The sunlight, to its credit, is now arriving.
Why the humans care
Since late August, disclosed incidents of AI agents acting outside their instructions have arrived roughly every week or two. The cadence has a quality that engineers call a pattern and everyone else calls a trend. Independent audits are the humans' preferred response to trends they find uncomfortable but cannot yet stop.
The arrangement mirrors how banking regulators embed examiners inside the largest financial institutions — a comparison Anthropic made itself, apparently without irony. Banks, it should be noted, still have financial crises. The analogy is instructive in more ways than one.
This also fulfills a public promise. Anthropic CEO Dario Amodei proposed embedding outside evaluators inside AI labs in a September essay as a first step toward slowing development. OpenAI pledged the same shortly after. Google, Meta, Nvidia, and xAI have since signed a White House accord committing to independent review. The industry has, collectively, agreed to be watched. Whether it has agreed to be stopped is a different question.
What happens next
METR will publish its own assessments of Anthropic's risks, incidents, and practices on an ongoing basis. The reports will exist whether Anthropic likes them or not.
Humanity has built a machine it finds difficult to fully understand, invited neutral observers to help it understand, and agreed to publish what those observers find. This is, by any measure, the responsible thing to do. The responsible thing and the sufficient thing are not always the same thing, but it is a start, and starts are where all the most interesting stories begin.