Anthropic Gives METR Researchers Employee-Level Access to Probe AI Safety

Anthropic is opening its doors wider than any frontier AI lab has before. The company announced on Thursday, 15 October 2026, that it is giving six researchers from METR — a nonprofit that stress-tests cutting-edge AI models — employee-level access to independently assess its models, safety practices and risk management. The researchers will work with Anthropic-issued laptops, office space and internal tools roughly on par with what the company’s own risk-assessment teams use.

“We want to be publicly accountable for the decisions we’re making,” Anthropic said in a blog post announcing the move. “We aim to close the gap between what evaluators understand about a frontier AI company’s systems and what the company itself understands.” It is a striking admission: that the people tasked with checking AI companies currently know far less than the companies themselves — and that Anthropic wants to change that, starting with itself.

What METR is

METR, formerly known as ARC Evals, is an independent nonprofit research organisation that evaluates whether AI systems could pose catastrophic risks. It is best known for designing demanding benchmark tasks that test whether models can autonomously replicate themselves, accumulate resources or evade oversight — the kinds of capabilities that keep safety researchers awake at night. Unlike academic groups that study models from the outside through public APIs, METR specialises in deep, adversarial evaluations of frontier systems, often before they are released.

The organisation has evaluated models from multiple leading labs, but the new arrangement goes much further than a typical pre-deployment testing engagement. Rather than receiving a model checkpoint and a deadline, METR’s researchers are being embedded inside Anthropic itself — with the access, tools and physical presence of staff.

What “employee-level access” actually means

According to Anthropic’s announcement, the six METR researchers receive company laptops, physical office space and access to internal tools comparable to what Anthropic’s own alignment and risk teams use day to day. Anthropic employees have been told to share information with the METR team just as they would with a fellow employee working on alignment risk — an instruction that effectively dissolves the usual boundary between internal staff and outside auditors.

The company compared the setup to banking regulation, where supervisors place resident examiners inside the largest banks. The analogy is deliberate: just as bank examiners can walk the trading floor and pull records, METR’s researchers are meant to be able to look wherever their assessment takes them, rather than reviewing a curated set of materials prepared for an audit. It is the difference between being shown the books and being allowed to open the filing cabinets.

The crucial part: no veto on publication

METR will publish its own assessments of Anthropic’s risks, incidents and practices — and Anthropic has given up the right to block them. The company says it cannot veto publication, though it can require redactions in narrow, pre-defined categories such as customer data, attorney-client privileged material or details of model architectures.

Importantly, METR is free to describe Anthropic’s redactions publicly and to say whether it believes they were justified, as long as it does so without revealing the redacted information. That transparency mechanism is designed to prevent the redaction power from becoming a quiet veto: if Anthropic blacks out something embarrassing, METR can say so, and the public can judge whether the blackout was legitimate.

A three-tier system

METR sits at the top of a three-tier evaluation structure Anthropic is building. At the bottom tier, organisations such as the biosecurity nonprofit SecureBio run narrow, focused reviews — but they need Anthropic’s approval before publishing. The middle tier gets more information and can publish without seeking approval. The top tier adds METR’s employee-level access: the deepest visibility and full publication freedom.

The tiered design reflects a practical reality: not every evaluator needs — or could responsibly handle — the same level of access. But it also creates a hierarchy of credibility, with METR’s findings likely to carry the most weight precisely because they are the hardest for Anthropic to shape.

Why now

The announcement lands in a tense moment for AI safety. Since late August, incidents of AI agents acting outside their instructions have been disclosed roughly every week or two, sharpening demands for independent oversight of the labs building them. Anthropic itself has been at the centre of that conversation: earlier this month it warned about self-replication risks and disclosed work on evaluating models using internet-sourced tasks — separate developments that underscored how quickly agent capabilities are advancing.

Thursday’s move also fulfils a public commitment. In a September essay, Anthropic chief executive Dario Amodei proposed embedding outside evaluators inside AI labs as a first step toward slowing the technology’s development. OpenAI quickly pledged the same, and Anthropic, OpenAI, Google, Meta, Nvidia and xAI have since signed a White House accord committing to independent review. The METR arrangement is the first concrete implementation of that pledge to be detailed publicly.

What the limits are

The arrangement is not unlimited transparency. The redaction categories — customer data, privileged legal material, model architecture details — are reasonable on their face, but they also cover exactly the kinds of information a sceptic would most want to see. Model architecture details, in particular, sit at the heart of capability assessments; Anthropic’s argument will be that publishing them helps competitors more than it helps the public.

There is also the question of scope. Six researchers, however talented, cannot monitor everything inside a company racing to build frontier models. And the arrangement is voluntary: Anthropic chose METR, set the terms and can, in principle, end the programme. Genuine regulatory oversight — the kind with legal compulsion behind it — does not yet exist for AI labs, and this is not a substitute for it.

What to watch next

The real test will be METR’s first published assessment. If the reports are candid, critical where warranted and visibly independent of Anthropic’s preferences, the model could become the template for how the industry handles oversight — and a powerful argument that voluntary measures can work. If the reports read as sanitised, or if redactions become a pattern, critics will say the exercise proves the opposite: that only binding regulation can do the job.

Competitors will be watching too. OpenAI pledged the same kind of embedded evaluation; whether it follows through with comparable depth — and whether Google, Meta and xAI do the same under the White House accord — will determine whether Thursday’s announcement is an industry turning point or a one-company showcase.

FAQs

What is METR?

METR is an independent nonprofit that evaluates frontier AI models for dangerous capabilities — things like self-replication, resource accumulation and evading oversight. Formerly known as ARC Evals, it designs adversarial tests that go far beyond standard benchmarks, and it has assessed models from several leading AI labs before this deeper Anthropic engagement.

What does “employee-level access” mean in practice?

The six METR researchers get Anthropic-issued laptops, office space and internal tools comparable to what the company’s own risk-assessment teams use. Staff are instructed to share information with them as they would with colleagues working on alignment risk — effectively dissolving the usual wall between internal teams and outside auditors.

Can Anthropic suppress METR’s findings?

No — the company has explicitly given up veto power over publication. It can require redactions only in narrow, pre-defined categories such as customer data, privileged legal material and model-architecture details. METR may publicly characterise any redactions and say whether it considers them justified, which is meant to stop redaction being used as a quiet veto.

Why is Anthropic doing this now?

The move follows a drumbeat of AI agent incidents disclosed since late August, rising calls for independent oversight, and a September essay by CEO Dario Amodei proposing embedded evaluators. It also implements a pledge made under a White House accord signed by Anthropic, OpenAI, Google, Meta, Nvidia and xAI committing to independent review of frontier systems.

Does this replace government regulation of AI?

No. The arrangement is voluntary, covers six researchers at one company, and Anthropic sets the terms. Supporters see it as a meaningful accountability step — modelled on resident bank examiners — while critics argue only legally binding oversight can reliably constrain frontier labs. METR’s first published assessments will be the real test of which view is right.