Anthropic said Thursday that six researchers from METR, the nonprofit that tests frontier models, now have employee-level access inside the company. They have Anthropic laptops, office space, and internal tools roughly on par with the lab's own risk teams. Their job is to assess models, incidents, and safety practices, and to publish what they find. Anthropic is not funding the work. TechCrunch said the lab compared the setup to resident bank examiners. It also said incidents of agents acting outside their instructions have been disclosed every week or two since late August, which is the run of news this access is answering.

TechCrunch's readout of the blog post and CNBC's account both describe a three-tier ladder. The bottom tier runs a narrow review and needs Anthropic's approval to publish. The middle tier sees more and can publish without that approval. The top tier is the employee-level access METR has. Anthropic said staff should share information with the METR group the way they would with a colleague on alignment risk. METR had already raised about $71 million as labs hired auditors. Thursday's change is not another funding round. It is a badge and a laptop.

Publication can be redacted. It cannot be blocked

Anthropic cannot veto a METR report. It can require redactions in narrow, pre-set categories: customer data, material that is attorney-client privileged, and information about model architectures. METR may say in public whether it thinks a redaction was justified, as long as the description does not reveal the redacted material. The assessments METR writes cover risks, incidents, and practices. Anthropic's line in the post was that it wants to be publicly accountable for its decisions, and that it wants evaluators to understand the systems about as well as the company does.

The arrangement follows a September essay by chief executive Dario Amodei, which proposed putting outside evaluators inside labs as a step toward slowing development. OpenAI said it would do the same. Late in September, a White House accord asked companies to partner with an independent auditor. Amodei signed, as did executives at Google, Meta, OpenAI, SpaceX, Nvidia, and xAI. METR is the only group on the top tier so far. CNBC also listed the accord's encouragement from President Donald Trump. Anthropic's Thursday line was that it is embedding evaluators voluntarily and that every frontier lab should.

SecureBio is on the ladder. Accenture is funded by the lab

SecureBio is the first-tier example. Its current job is the chemical and biological weapons section of the risk report Anthropic released in August. Broader access and publication rights could come later. CNBC said Anthropic is also in talks with Hugging Face and has not said which tier that would be.

Accenture was named as a partner days after the September essay. Anthropic said it would pay Accenture directly because of the urgency, and CNBC noted that the payment drew criticism on social media from people who questioned the independence. The METR terms are the contrast: no Anthropic funding, and no veto. Anthropic tested a worm that copies model weights onto cheap hardware. That was the lab checking its own systems. The METR team is there to check the lab. OpenAI has scrapped a model release over safety concerns, which is the other shape this work can take when the evaluator and the lab are the same company.

The tiers still have no outside standard

Anthropic has said there are no settled rules yet for what an embedded evaluator should see, how the findings should be reported, or who should pay. Its longer-term preference, stated with the Accenture deal, is pooled or government money. Neither exists. Until it does, the Thursday structure is the one on paper: METR at the top, unpaid, able to publish, and SecureBio on a narrower brief it cannot publish until Anthropic says so.