Story
Oct 18, 2026
Anthropic Opens Its Doors to AI Watchdogs—But Keeps the Redaction Key
Anthropic casts its new arrangement with METR as a practical route to public accountability without exposing sensitive systems, while outside observers see a significant transparency step tempered by the company’s continuing control over what must stay confidential.
Last month, Anthropic CEO Dario Amodei called for frontier AI companies to slow down and bring outside evaluators inside their operations. The company also signed the White House Accord on Super Intelligence, joining other major technology groups in pledging to work with independent auditors.1
On Thursday, Anthropic made good on that commitment by bringing six researchers from the nonprofit Model Evaluation and Threat Research, or METR, into its offices. The team has company hardware, workspace and internal-tool access “roughly on par” with Anthropic’s own risk-assessment staff.1 Anthropic says the point is to close the information gap between companies building powerful models and outsiders asked to judge their risks.
The lab’s case is straightforward: stronger models require deeper scrutiny, but releasing every detail of training, computing and safeguards could aid malicious actors or reveal commercially sensitive information. It compares the arrangement to resident examiners placed inside major banks to spot dangerous risk-taking before it harms the public.1
METR, however, is not simply an internal team with a new badge. Anthropic says it cannot veto the group’s reports on risk levels, incidents and practices. It can demand redactions only for defined categories, including customer data, attorney-client material and model-architecture details; METR may publicly say whether it believes those redactions were justified.1
That caveat defines the tension around the experiment. Reporting on the announcement noted critics’ concern that evaluator arrangements can create conflicts of interest and may fall short of government regulation.2 Anthropic says it is not funding METR, whose work is backed by philanthropic and government sources, a structure intended to protect its independence.2
The company is also building a three-tier system. SecureBio, reviewing chemical and biological weapon risks, currently has narrower first-tier access and publication rights, while METR alone occupies the top tier. Anthropic is still discussing partnerships with organizations including Hugging Face and working toward a broader arrangement with Accenture.3 The result is a voluntary oversight model with unprecedented access—but one whose real credibility will rest on what METR is able to publish.