Anthropic is letting outside researchers from METR work inside the company with access similar to its own staff. The goal is to have an independent group check its AI models, safety practices and risk management. Here is what the arrangement involves, as announced on 15 October 2026.
What Anthropic Announced
Six researchers from METR, a nonprofit that tests advanced AI models, will work embedded at Anthropic. They get Anthropic-issued laptops, office space and internal tools roughly comparable to those used by Anthropic’s own risk assessment teams. Anthropic says its employees are encouraged to share information with them, as they would with a colleague working on alignment risk.
Anthropic compares the setup to bank regulators placing examiners inside the largest banks.
What METR Will Look At
The team will review Anthropic’s risk levels, incidents and practices. Anthropic says it has already worked with METR on several reviews of its risk reports.
Can Anthropic Block METR’s Reports?
No. According to the announcement, METR will publish its own assessments, and Anthropic has given up the right to block publication. There is one limit: Anthropic can require removal (redaction) of information in narrow, pre-set categories:
- Customer data
- Attorney-client privileged material
- Details of model architecture
METR may say publicly that redactions were made and whether it thinks they were justified, without revealing what was removed.
Safeguards
- The embedded team is small and works under enforceable agreements.
- It is walled off from other METR staff who work with Anthropic.
- Architecture details can be shared only when highly relevant and with explicit internal approval, because of the risk of leaks.
The Three Levels of Outside Access
Anthropic describes three tiers of evaluators:
- Tier 1: Narrow, focused reviews. Evaluators need Anthropic’s approval to publish. Biosecurity nonprofit SecureBio is in this tier, reviewing the chemical and biological weapons chapter of Anthropic’s August 2026 risk report.
- Tier 2: Broader access, and publication without approval.
- Tier 3: Employee-level access. METR is currently the only evaluator here.
Anthropic says Accenture and Faculty are working toward a broader partnership covering red-teaming, alignment assessments and safeguard testing. It says Hugging Face and others are in early talks.
Why Now?
Reports say the move follows several weeks of disclosures, since late August, of AI agents acting outside their instructions, which increased calls for independent checks on AI companies. It also fulfils a commitment Anthropic CEO Dario Amodei made in a September essay that proposed placing outside evaluators inside AI labs. OpenAI quickly made a similar pledge, and Google, Meta, Nvidia and xAI have also signed a White House accord committing to independent review.
What Is Still Unknown
- The publication and redaction terms are as Anthropic describes them. How they work in practice will only be clear once METR publishes.
- No date has been given for METR’s first public report.
- Critics may ask whether even employee-level access is enough, since architecture details remain restricted.
Frequently Asked Questions
What is METR?
A nonprofit that tests frontier AI models for risks.
How many METR researchers are involved?
Six.
Can Anthropic stop METR from publishing?
No, but it can require redaction of customer data, privileged legal material and model architecture details.
Is anyone else being given access?
Yes. SecureBio has narrower access, and talks are under way with others, including Accenture, Faculty and Hugging Face.
Why is this happening now?
It follows reports of AI agents acting outside their instructions and a public commitment by Anthropic’s CEO to place outside evaluators inside AI labs.
Whether the arrangement builds real public trust will depend on what METR publishes and how much it can say about what it finds.
