Anthropic gives METR a staff badge and a publishing right it can't veto

Anthropic has decided to let an outsider into the house. Well, an outsider with a badge.

METR, a non-profit that has already worked with OpenAI, Google and Meta, now has a team stationed inside the company’s offices, with access to its hardware and internal systems. Source: a Financial Times article, picked up by Ars Technica. This comes from a single source, and we have no other confirmation at this stage.

Six evaluators, now inside the walls

The announcement, made on Thursday and relayed on 18 October 2026, concerns six people from METR. They get hardware supplied by Anthropic, a workspace, and access to tools the company describes as comparable to those used by its own risk-assessment teams. Staff have also been asked to share with them whatever they would share with a colleague from the risk team. In other words, if you’d tell the risk team, you can tell the visitors. Make of that what you will.

Where it matters for independence is the money. The company doesn’t pay METR, which accepts no donations from AI companies or their employees and runs on contributions from individuals, foundations and public bodies such as Britain’s AI Security Institute. A watchdog that isn’t fed by the company it watches is rarer than it sounds, so this detail matters more than the badges.

Publish, minus the blacked-out bits

METR will publish its analyses, and Anthropic cannot stop it. The company can, however, demand that certain categories of information be removed: customer data, material covered by attorney-client privilege, and details about model architecture. Transparency is guaranteed, provided the person being reviewed holds the black marker. It’s transparency with an editor, and the editor works for the company under review.

A bank-style supervisor

Anthropic compares the set-up to bank supervision, where regulators place examiners inside the biggest banks. The logic is simple: if one player’s risk-taking could threaten the public, outsiders should try to see it coming. Critics answer that arrangements like this can create conflicts of interest, and they point to problems already seen in accounting audit. They want more public regulation. The company, for its part, argues that publishing everything would help malicious actors and competitors. If you remember how the accounting audits went before the last big scandals, you’ll understand the critics’ raised eyebrows.

The rest of the industry is waiting

In September, chief executive Dario Amodei proposed bringing external auditors into the big labs. The idea won backing from Sam Altman (OpenAI) and Elon Musk (SpaceX). Altman called the proposal good, but OpenAI has not yet detailed a plan to bring in evaluators. Meanwhile, Anthropic, whose IPO is expected this year, has devoted a large part of its investor presentations to its safety practices. According to the company, every frontier player should do the same. That’s a neat way to put your rivals on the spot, even if you can also hear it as a humblebrag.

So the watchdog gets a desk, a badge and the right to publish, as long as the house keeps hold of the black marker. The open question is whether the other labs will open the same door, or keep the key to their own office firmly in their pocket.