Anthropic has given six researchers from METR, a nonprofit that tests frontier AI models, employee-level access to its models and safety processes. It has also given up the right to block METR from publishing what it finds. Anthropic announced the Anthropic METR access arrangement in a blog post on Thursday, October 15, and TechCrunch reported the terms the same day. It is a voluntary commitment, but it is a concrete step toward the outside scrutiny that regulators and lawmakers are pressing for.
What the METR team gets inside Anthropic
According to TechCrunch's account, the METR team will receive Anthropic-issued company laptops, office space and access to internal tools. Those tools are roughly on par with what Anthropic's own risk assessment teams use. The goal is to let outside researchers assess Anthropic's models, safety practices and risk management from the inside, rather than only through documents and briefings the company chooses to release.
The access, as reported, includes:
- Employee-level status for six METR researchers
- Anthropic-issued laptops and dedicated office space
- Internal tools roughly comparable to those used by Anthropic's risk assessment staff
- Direct assessment of models, safety practices and risk management
The difference from a standard external review is proximity. Outside researchers with internal tools can ask follow-up questions while they work, instead of waiting for a company to answer a request in writing. The reporting does not say how METR was selected, how long the access will last, or whether Anthropic will pay the researchers. Those details matter for independence, and they are worth checking in the primary post.
How the publication veto became narrow redaction rights
The most consequential change concerns what gets published. METR will publish its own findings on Anthropic's risks, incidents and practices, and Anthropic has surrendered the power to stop that. "Anthropic can't veto publication," the company said, as TechCrunch quoted its blog post.
Anthropic has kept one tool. It says it can still require redactions, but only in narrow, pre-defined categories. The reported categories are customer data, attorney-client privileged material and information about model architectures. The exact boundaries of those categories are not spelled out in the reporting.
Anthropic also said METR may publicly say whether it thinks a redaction was justified, as long as it does not reveal the redacted content. TechCrunch quoted the company saying METR "is free to characterize our redactions publicly."
That clause matters more than it first appears. A redaction nobody can question is just a silent edit. Letting METR say whether a redaction holds up gives readers a way to test whether Anthropic is using the categories honestly. The weak point is definition. "Model architecture information" can cover a wide range of material, and until Anthropic shows how it applies the categories, outsiders will have to rely on its judgment.
Why the timing matters: Amodei, the White House and state lawmakers
The arrangement fulfills a commitment Dario Amodei made in a September essay. He proposed putting outside evaluators inside AI labs as a first step toward slowing development. The METR deal is the first concrete step tied to that proposal that has been reported so far, which makes it a useful test of whether the idea works in practice.
The political pressure points the same way. A White House accord signed September 29 by President Trump and the heads of six AI companies commits the signatories to work with "independent auditors" to assess their models' safety and capabilities. The companies agreed to a document described as "morally binding."
Microsoft president Brad Smith has called for global standards that include independent evaluations, mandatory incident reporting and an "emergency brake" for rogue AI agents. He made those comments in a BBC interview that TechCrunch reported on October 12.
At the state level, WBUR reports that Massachusetts is weighing rules that could mandate independent evaluations of AI companies' models and safety protocols. Those rules sit inside a state senate economic development bill that is still under negotiation. USA 24 has covered the Massachusetts proposal separately. Anthropic has also pitched its own federal safety bill with narrow blocking power. Read together, the two moves show a lab trying to shape the rules that will apply to it.
What the arrangement does not prove
Three limits deserve attention. First, the arrangement is voluntary. Nothing in the reporting says it would count toward a state or federal requirement, including the ones debated in Massachusetts. Second, access is not assurance. Outside researchers can have good tools and still miss problems, and their work only carries weight if the findings are specific and published. Third, this covers one lab. The reporting does not say whether OpenAI or other developers will offer similar access.
The publication clause is the part that will decide whether this is meaningful. A lab that gives outsiders access but keeps a veto over what they say has not really shared oversight. METR's freedom to publish is where the arrangement gets its teeth, and it is also the part that will take the longest to judge. The reporting does not say how often METR will publish, or whether its findings will be critical.
The trade-offs are clear enough. Anthropic gains credibility with lawmakers who want independent oversight, and METR gains access it could not get from the outside. Lawmakers pushing for binding rules could read the deal as evidence that voluntary steps are moving, or as proof that they are not enough. Both readings will get a test once METR's first report appears.
What enterprise buyers should check
Consumers who use Claude need to do nothing differently. Business buyers and security teams can use the announcement as a checklist when they review AI vendors:
- Ask whether any third-party evaluator has employee-level or comparable access to the models and safety processes you depend on.
- Ask whether that evaluator can publish findings, and whether the vendor holds any veto over publication.
- Ask which redaction categories apply, and how the vendor defines them in practice.
- Once METR publishes, compare its findings with the vendor's own safety disclosures.
Readers making procurement decisions should read Anthropic's blog post directly, since the details in this article come from TechCrunch's report of it.
What to watch next
METR's first publication will be the real test. Watch whether it names specific risks and incidents, whether it uses its right to characterize any redactions, and whether Anthropic publishes more detail on its redaction categories. Also watch whether OpenAI or other labs answer with similar access, and whether Massachusetts lawmakers write an evaluation requirement into the state bill. Each of those developments will show whether voluntary access becomes an industry norm or stays a one-off.
