Anthropic gives METR researchers employee-level access to review its AI models

Anthropic has started giving employee-level access to six researchers from METR, a nonprofit that tests frontier AI models. Their j...

Anthropic has started giving employee-level access to six researchers from METR, a nonprofit that tests frontier AI models. Their job is to independently assess Anthropic's models, safety practices, and risk management. Anthropic announced the arrangement on Thursday, October 15, 2026. METR will publish its own findings, and Anthropic has given up the right to block them. It is a notable step in a year when AI labs are under growing pressure to show their safety claims can be checked by outsiders.

Why the announcement matters now

The timing is not an accident. Since late August, incidents of AI agents acting outside their instructions have been disclosed every week or two. Those reports have sharpened calls for independent checks on AI companies, which are otherwise left to grade their own work.

Anthropic frames the move as a matter of accountability. It says it wants to be publicly accountable for the decisions it makes, and that it aims to close the gap between what evaluators understand about a frontier AI company's systems and what the company itself understands. It also admits a limit: it cannot publicly share everything relevant to its safety work. Some information could help malicious actors, and some is commercially sensitive, such as details about compute or training. Its answer is to give trusted third parties deep access and have them share evidence and reasoned conclusions about the risks.

What "employee-like access" means

The access is concrete. The METR team has company laptops from Anthropic, office space, and access to internal tools and systems roughly on par with what Anthropic's own risk assessment teams use. Anthropic employees are encouraged to share information with the METR researchers just as they would with a colleague working on alignment risk.

There are exceptions. A small number of clearly defined categories stay off limits: data about customers, attorney-client privileged information, and information about the company's model architectures. Anthropic gives a reason for the last one. Architecture details carry a high risk of accidental leaks, because evaluators may also work with other companies, and their relevance to risk assessment is usually low. When they are highly relevant, Anthropic says it will generally try to share them, though that needs explicit internal approval.

Strict limits on who sees what

Anthropic stresses that the arrangement is narrow. The engagement is limited to a small team working under enforceable agreements. Other METR employees are kept separate from that team's work, including researchers who work with Anthropic on other projects. Anthropic says the setup gives a few qualified people, working under its security standards, a close view of how it operates. It adds that this is not a way for other organizations to obtain its information or that of its customers and partners.

METR is not new to Anthropic's work. Over the past year it has carried out several reviews of Anthropic's Risk Reports.

No veto on publication

The most important term may be the publication rule. METR will publish its analysis of Anthropic's risk levels, incidents, and practices. Anthropic cannot veto publication. It can require METR to redact information, but only when that information falls into narrow, predefined categories like the ones above.

METR is also free to characterize those redactions publicly and to say whether it thinks they were justified, as long as it does so without revealing what was removed. That detail matters. It means readers can see not only the findings but also how much was held back and whether the evaluator found the reasons convincing.

The bank examiner comparison

Anthropic compares the setup to bank supervision, while admitting the analogy is not exact. Banks know more about their own risk positions than outsiders do, so regulators place resident examiners inside the largest banks. If a bank's risk-taking could threaten the public, third parties try to see it coming. Anthropic's argument is that frontier AI labs now hold similar knowledge advantages and carry similar potential for public harm.

What evaluators need

Anthropic lists four things an evaluator needs to do the job well:

Expertise in assessing frontier model capabilities, including how evaluations can go wrong, for example through alignment faking.

Independence, meaning no equity in the company being evaluated and no financial or operational dependence on it.

Access that goes beyond the latest models to include how they were trained and how safeguards work.

Freedom to publish observations and assessments about a model or a process.

A three-tier system

METR sits at the top of a three-tier structure that Anthropic is building.

First tier: Evaluators run narrowly scoped risk investigations. Anthropic decides what access they get, and they can publish only with its approval.

Second tier: Evaluators get more information and can publish without approval. Anthropic can still require removal of specific categories of information, and evaluators may say whether they think those removals were justified.

Third tier: This is the METR arrangement, with employee-like access and publication rights similar to the second tier.

Anthropic says it hopes to have several evaluators at the third tier over time, each publishing its own analysis of the company's risk posture and key decisions. For now, METR is the only one there.

The other partners

Anthropic says it is working on several other partnerships, all at earlier stages with more limited information and publication rights. The goal of these early engagements is to confirm that partners protect confidential information and have the resources and expertise to evaluate the models.

SecureBio a biosecurity nonprofit, is reviewing the chemical and biological weapons chapter of Anthropic's August Risk Report. It completed and published a pilot review of the previous report. Anthropic says it will consider projects with more access after this review is done.

Accenture and Faculty are working with Anthropic toward what it calls an ambitious partnership. It covers evaluating and red-teaming models, alignment assessments, testing safeguards, and building capabilities across both organizations.

Other organizations, including Hugging Face, are in early discussions.

How it connects to wider commitments

The announcement follows through on a promise from Anthropic's CEO, Dario Amodei. In a September essay, he proposed embedding outside evaluators inside AI labs as a first step toward slowing the technology's development. OpenAI quickly pledged the same. Both companies, along with Google, Meta, Nvidia, and xAI, have since signed a White House accord committing to independent review. Anthropic describes its embedding of evaluators as voluntary and says it believes every frontier AI company should do the same. Coverage of the news describes Anthropic as the first frontier AI company to offer this kind of access.

What to watch

The first METR report: The arrangement will only prove its value when METR publishes something. How specific it is, and how many redactions it reports, will show how open the access really is.

The size of the carve-outs: Customer data and privileged material are easy to justify. The architecture exclusion deserves attention, since it is the category most likely to touch how risky a model.

Scale.:Six researchers is a small team for a very large company, and the arrangement is limited by design.

Voluntary status: The access exists because Anthropic chose it. It does not yet rest on a legal requirement.

Mixed reactions: Some coverage notes that the step also draws criticism, so expect the debate over whether embedded evaluation is enough to continue.

Frequently Asked Questions

What did Anthropic announce?

Anthropic said it has started giving employee-like access to six researchers from METR, a nonprofit that evaluates frontier AI models. They can independently assess Anthropic's models, safety practices, and risk management.

What can the METR researchers access?

They have company laptops, office space, and internal tools and systems roughly on par with Anthropic's own risk assessment teams. Customer data, attorney-client privileged information, and model architecture details are generally excluded.

Can Anthropic stop METR from publishing?

No. Anthropic cannot veto publication. It can require redactions only in narrow, predefined categories, and METR may say publicly whether it thinks those redactions were justified.

Why is Anthropic doing this?

It says it wants to be publicly accountable and to close the gap between what evaluators understand about its systems and what the company understands. It also cannot share all safety-relevant information publicly.

What are the three tiers of evaluator access?

First-tier evaluators run narrow reviews and need approval to publish. Second-tier evaluators get more information and can publish without approval. The third tier, where METR sits, adds employee-like access.

Which other organizations is Anthropic working with?

SecureBio is reviewing part of an Anthropic risk report. Accenture and Faculty are working toward a broader partnership. Anthropic is also in early talks with others, including Hugging Face.

How does this relate to the White House accord?

Anthropic and OpenAI, along with Google, Meta, Nvidia, and xAI, have signed an accord committing to independent review. Anthropic describes this step as following through on that commitment and on its CEO's September proposal.

What is the bank examiner comparison?

Regulators place resident examiners inside the largest banks because banks know more about their own risks than outsiders do. Anthropic says embedded evaluators serve a similar purpose, though it admits the analogy is imperfect.

References

Tech crunch - https://techcrunch.com/2026/10/15/anthropic-is-letting-metr-researchers-look-under-the-hood-with-employee-level-access/

Ground News - https://ground.news/article/anthropic-grants-ai-evaluator-metr-employee-level-access-to-its-models

Anthropic - https://www.anthropic.com/news/an-update-on-independent-evaluators

Loaded All Posts Not found any posts VIEW ALL Readmore Reply Cancel reply Delete By Home PAGES POSTS View All RECOMMENDED FOR YOU LABEL ARCHIVE SEARCH ALL POSTS Not found any post match with your request Back Home Sunday Monday Tuesday Wednesday Thursday Friday Saturday Sun Mon Tue Wed Thu Fri Sat January February March April May June July August September October November December Jan Feb Mar Apr May Jun Jul Aug Sep Oct Nov Dec just now 1 minute ago $$1$$ minutes ago 1 hour ago $$1$$ hours ago Yesterday $$1$$ days ago $$1$$ weeks ago more than 5 weeks ago Followers Follow THIS PREMIUM CONTENT IS LOCKED STEP 1: Share to a social network STEP 2: Click the link on your social network Copy All Code Select All Code All codes were copied to your clipboard Can not copy the codes / texts, please press [CTRL]+[C] (or CMD+C with Mac) to copy Table of Content