Who actually checks whether a powerful AI model is safe before it lands on your phone, your laptop, or your company’s helpdesk? For most of 2026, the honest answer was mostly the company that built it. That changed on October 15, 2026, when Anthropic said it is giving an outside nonprofit the kind of deep, ongoing access to its systems that usually only its own staff get.

This matters even if you never read an AI safety report yourself. It is part of a bigger shift toward trustworthy AI, where AI safety evaluators are being given real power to check a company’s work instead of just taking its word for it. Here is what actually happened, why it is a bigger deal than the headline suggests, and what it means for anyone who uses AI tools day to day.

What Are AI Safety Evaluators, Really?

Think of an AI safety evaluator as an outside inspector. A restaurant does not get to grade its own health inspection, and AI labs are slowly being asked to accept a similar idea. An evaluator’s job is to test a model for real risks: can it be tricked into helping with something dangerous, does it behave consistently under pressure, does it hide information from the people checking it. If you are new to AI concepts in general, our simple explanation of what AI actually is is a good place to start before this one.

Until recently, most of this checking happened inside the AI companies themselves, using their own staff and their own rules. Outsiders could ask questions, but they rarely got real access to the systems or the right to publish what they found without approval first.

What Anthropic Just Announced

On October 15, 2026, Anthropic published an update saying it had started giving a six person team from METR, a nonprofit that evaluates frontier AI models, something close to employee level access. According to Anthropic’s own announcement, the METR team gets company laptops, office space, and internal tools similar to what Anthropic’s own risk assessment teams use. Staff are encouraged to share information with METR the way they would with an internal colleague working on AI risk.

The part that really stands out: METR can publish its own findings about Anthropic’s risks and safety incidents, and Anthropic has given up the right to block that publication. The only thing Anthropic can still do is ask for narrow redactions, things like customer data, legally privileged material, or details of exactly how the model works internally. Even then, METR is allowed to say that something was redacted and whether it thinks the redaction was fair.

Anthropic is not stopping at METR either. A biosecurity focused nonprofit called SecureBio is separately reviewing the chemical and biological weapons section of Anthropic’s safety reporting, and Anthropic is working toward a broader partnership with Accenture’s AI division, Faculty, covering model testing and red teaming more generally.

Why Would a Company Give Up That Much Control?

From my own experience working with websites, online tools, and cybersecurity projects, I know how tempting it is to only show people the parts of a system that make you look good. Handing an outsider real access and the right to publish the unflattering parts too is a very different kind of choice. The real tip here: when a company gives up its own veto over bad news about itself, that usually tells you more than any press release does.

Anthropic frames this as following CEO Dario Amodei’s September 2026 proposal that outside evaluators should sit inside frontier AI labs, similar to how bank examiners sit inside major banks, as an early step toward slowing down reckless AI development. It also lines up with a bigger moment: in late September 2026, Anthropic, OpenAI, Google, Meta, Nvidia, and xAI signed a voluntary White House accord on AI safety, agreeing in principle to independent review of their most powerful models. It is the same broader trust and safety push behind Anthropic’s recent AI cybersecurity models announcement and its report on rising AI linked cyberattacks.

Not everyone is convinced this solves the real problem. Critics point out that Anthropic is still paying the evaluators, which makes this closer to a company hiring its own referee than a fully independent regulator checking in uninvited. Anthropic’s response is that outside evaluators do not reduce its own accountability, they make that accountability easier to verify from the outside, and that the safety of its models remains Anthropic’s responsibility either way.

What This Means for You

You will probably never read METR’s full reports, and that is fine. What matters is that there is now a paper trail: a named outside group with real access, real publishing rights, and a track record you can look up for yourself. If you use Claude, ChatGPT, or Gemini at work or for study, this kind of oversight is part of why these tools keep getting examined for failure modes like deception or unsafe autonomous behavior. Anthropic’s own models keep shipping fast while this oversight work runs alongside them, including recent releases like Claude Haiku 5.5.

It is also a preview of where AI regulation is probably heading next. Expect more “who checks the checker” stories like this one over the coming year, as governments and companies work out what independent AI oversight should actually look like in practice, not just on paper.

Common Questions

Is METR part of Anthropic?
No. METR is an independent nonprofit that evaluates AI models from more than one company, not just Anthropic.

Can Anthropic block what METR publishes?
No. Under this arrangement Anthropic gave up the right to block publication, though it can still request limited redactions for things like customer data or model architecture details.

Does this mean Claude is now proven safe?
No. It means an outside group has deeper access to check Anthropic’s safety claims. Evaluation reduces uncertainty, it does not hand out a permanent guarantee.

Are other AI companies doing something similar?
OpenAI has made a comparable commitment, and Anthropic, OpenAI, Google, Meta, Nvidia, and xAI all signed a voluntary White House accord in September 2026 agreeing to independent review in principle.

Final Takeaway

AI safety used to mean companies mostly grading their own homework. Anthropic’s October 2026 arrangement with METR is a real, if imperfect, step toward letting outsiders check that work instead, and it is worth watching whether OpenAI, Google, and the rest follow with the same level of real access rather than just the same language. For now, treat it as a sign that trustworthy AI is becoming something companies have to show, not just something they get to say. You can read Anthropic’s own account in its October 15 update on independent evaluators, with additional reporting from TechCrunch and background on the evaluators themselves at METR’s own site.