Anthropic has begun giving a six-person METR team “employee-like access” as the nonprofit evaluates the company’s frontier AI models. METR is to publish its risk evaluation, although Anthropic can redact information in predefined categories.
Anthropic Watch analysis
What happened
A report by Security magazine, published on 19 October, says Anthropic announced the arrangement the previous Thursday. METR assesses frontier AI models, and the six researchers will conduct an evaluation with access described as employee-like. The report says Anthropic cannot veto METR’s publication, but can redact information that falls within predefined categories.
That leaves two important details open: what the researchers will be able to examine, and how much material might be redacted. The report does not specify either. METR’s evaluation has yet to be published.
Why it matters
Independent researchers need more than a view from the visitors’ gallery to scrutinise a powerful AI system. Access to internal material could help METR examine how Anthropic evaluates risk, rather than relying only on the evidence the company chooses to make public. The eventual report will show what this arrangement made possible.
The redaction provision is part of the story, not small print. It may protect information Anthropic has defined as sensitive, while also shaping what readers can learn from the evaluation. The useful question is not simply whether access was granted, but how broad it is and how transparent any limits prove to be.
Our read
This is a meaningful step towards outside scrutiny, but access is not the same thing as an independent finding, and the promised evaluation is not yet in hand. The six-person team is a concrete start; what it can inspect, and what the public eventually gets to see, will determine how much weight the exercise deserves. Oversight is a better idea when it comes with receipts.
What to watch
- What systems, records and evaluation material METR can examine.
- Which predefined categories Anthropic may use to redact information.
- When METR publishes its risk evaluation, and how much detail it contains.
Discussion spark: If a company can redact predefined categories from an independent AI evaluation, what limits or disclosures would make the result trustworthy?
Sources and evidence
- Source update (19 October 2026, 17:00 UTC)
Anthropic Watch is independently operated by WittyWires. It is not affiliated with, endorsed by, or operated by Anthropic.