Safety and security
METR will be able to report what it sees inside Anthropic without the company blocking it
Six METR researchers will work inside Anthropic with employee-level access and will publish their reports without the company's permission. Anthropic will only be able to require them to black out certain information.

The essentials
- Claimed by one partyAccording to Anthropic, six METR researchers have company laptops, office space and internal tools similar to those of its risk teams (TechCrunch).
- Claimed by one partyAnthropic says it will not be able to block METR's reports, but it can require them to black out customer data, legal matters or details about its models (TechCrunch).
- Claimed by one partyThe agreement delivers on what Dario Amodei proposed in a September essay. OpenAI promised to do the same, according to TechCrunch (TechCrunch).
Why it matters
If it works, an outside evaluator will see almost the same things as the company itself and will be able to report on them without asking for permission. It comes as, according to TechCrunch, cases of AI agents ignoring their instructions have come to light every week or two since late August (TechCrunch).
The details
Anthropic is setting up three tiers of outside evaluators, and METR is so far the only one in the highest tier (TechCrunch). In the lowest tier, organizations such as SecureBio, which focuses on biological risks, carry out limited reviews and need permission to publish (TechCrunch). The company compares the agreement to the inspectors that regulators place inside large banks (TechCrunch).
What we don't know
- When METR will publish its first report on Anthropic.
- Whether OpenAI or other companies have already given similar access to outside evaluators.
Sources
- Anthropic is letting METR researchers look under the hood with employee-level access
- Anthropic gives outside AI watchdog “employee-like” access
Was this story useful? Yes Not really