Anthropic lets outsiders judge its AI — and can't stop them

Anthropic just did something no other frontier AI lab has done: it let outsiders inside the building, gave them employee-level access to its systems, and told them they can publish whatever they find — without asking permission first.

On October 15, 2026 (UTC), the company announced that six researchers from METR, a nonprofit AI evaluation group, have moved into Anthropic's San Francisco offices with company hardware, internal tool access, and a mandate to publish risk assessments that Anthropic cannot veto. It's the first concrete implementation of a pledge that Dario Amodei, Sam Altman, and Elon Musk jointly endorsed back in September — and it lands barely a week after Reuters reported that both Anthropic and OpenAI were privately negotiating to retain editorial control over exactly these kinds of reports.

What "employee-like access" actually means

The six METR evaluators get access roughly on par with Anthropic's own risk-assessment team. They can use internal tools, review incident data, and talk to employees about alignment work. Three categories are off-limits: customer data, attorney-client privileged information, and model architecture details. Everything else is fair game.

The critical detail is publication rights. METR will release its analysis of Anthropic's risk levels, incidents, and safety practices. Anthropic can require redactions only for those narrow pre-defined categories — and METR is free to publicly say whether it thinks a redaction was justified. That's a meaningful difference from the standard NDA-protected audit, where the company being audited controls what sees daylight.

Anthropic is also building a three-tier system for evaluators. Tier one gets narrow investigations and needs approval to publish. Tier two can publish without approval but faces broader redaction rights. METR is at tier three — employee-like access plus publication freedom. SecureBio is currently at tier one reviewing the biosecurity chapter of Anthropic's August Risk Report. Accenture and Faculty are in early-stage partnerships. Hugging Face is in early discussions.

The gap between pledge and practice

This announcement matters more because of what happened between the pledge and the implementation. On September 12, Amodei published "We Must Pace the Frontier," arguing for embedded independent evaluators inside frontier labs. Altman endorsed it within a day. The White House Accord on Super Intelligence, signed in late September, made independent auditing one of four "morally binding" commitments.

Then Reuters reported on October 13 (UTC-4) that as recently as October 5 (UTC-4) — less than three weeks after the public commitments — both Anthropic and OpenAI were privately pushing for strict editorial control over what safety evaluators could release. Anthropic called the characterization inaccurate for "all of its discussions," which implicitly conceded it applied to some. OpenAI acknowledged removing details from at least one evaluation while letting reviewers flag redactions.

Fortune piled on October 12 (UTC-4), reporting that OpenAI had fired three safety researchers — including Tomek Korbak, who served as a point of contact for an independent evaluator investigating the Hugging Face incident — shortly after signing the White House Accord. The employees dispute the firing rationale.

Anthropic's October 15 announcement is, in effect, the answer to that reporting. It's the company saying: we're not just talking about independence — we're giving up the veto.

Why it matters — and why it's not enough

The bank-supervision analogy Anthropic uses is instructive. Regulators place resident examiners inside the largest banks because banks know more about their own risk than outsiders ever could. The same logic applies to frontier AI: a company training models with capabilities it doesn't fully understand needs someone inside who can ask uncomfortable questions.

But an October arXiv paper, "Who Can Say No? Control, Accountability and the Pace of Frontier AI," argues that Anthropic's model lacks three things that make bank supervision work. Bank examiners aren't paid by the bank — if they were, the bank could stop paying. They have an escalation ladder: question, warning, order, with everyone knowing the next rung. And at the top, they have action power — freeze a product, block a dividend, require more capital. METR has a desk in the building and a publication right. It has none of the three.

That's not a reason to dismiss the move. A published negative assessment from an embedded evaluator would be impossible to spin away, especially in the middle of an IPO roadshow. The reputational cost alone creates accountability that didn't exist before. But it's accountability through embarrassment, not through enforcement.

The OpenAI question

Anthropic's move puts pressure on OpenAI, which made the same September pledge but hasn't announced an equivalent embedded program. OpenAI published principles for third-party evaluation on September 22 and fired safety researchers in early October. If METR publishes a critical assessment of Anthropic in the coming months — and OpenAI has nothing comparable to point to — the contrast will be stark.

It also matters for Anthropic's IPO. The company is reportedly targeting a $1.5 trillion valuation, down from an earlier $2 trillion expectation, with a roadshow possibly starting the week of November 9. Independent evaluators publishing unvetted risk assessments during a quiet period would be unprecedented. Investors will get transparency they didn't ask for — and that may be exactly the point.

What to watch

Three things will tell us whether this is a genuine shift or a well-timed PR move ahead of the IPO. First, does METR actually publish something critical in the next 90 days, and how does Anthropic respond? Second, does OpenAI announce its own embedded evaluator program, or does the gap widen? Third, does the White House Super Intelligence Task Force — which held its first meeting October 15 and is drafting protocols for AI labs to report breaches — turn voluntary embedding into a regulatory requirement?

Anthropic has placed a bet that radical transparency is worth the risk. The next few months will reveal whether the bet pays off — or whether the first unvetted negative assessment makes every other frontier lab dig in.