Anthropic gives outside auditors employee-level access
Six METR researchers just got Anthropic laptops and internal tool access to audit models. Anthropic cannot veto their public reports.
Policy · Source: TechCrunch
What happened
Anthropic just handed the keys to outside auditors. Six researchers from the nonprofit METR now have company laptops, office space, and access to internal risk assessment tools. Anthropic staff are told to treat them like regular colleagues. The goal is to independently assess Anthropic models and safety practices.
Anthropic gave up the right to block METR from publishing its findings. The company can only demand narrow redactions for things like customer data, legal secrets, or model architecture. If METR thinks a redaction is unfair, they are free to say so publicly. Anthropic compares this setup to regulators placing resident examiners inside major banks.
This is the top level of a new three-tier access system. Evaluators like SecureBio run narrow checks at the bottom and need permission to publish. METR sits at the top with full employee-level access and no publication veto. Anthropic is also talking with Hugging Face and Accenture AI business Faculty for future audits.
Key facts
- 6 — Number of METR researchers given employee-level access to Anthropic
- 3 — Tiers of access Anthropic is building for outside evaluators
Why it matters
Trust in AI is shifting from marketing claims to proven audits. Since late August, reports of AI agents ignoring instructions have spiked. Builders relying on frontier models need to know the underlying infrastructure is safe. Having independent auditors inside the lab means developers get a clearer picture of actual model risks, not just the sanitized corporate version.
This forces the rest of the industry to open up. Reuters noted that AI labs previously fought to control what reviewers could publish to protect trade secrets. By letting METR publish without a veto, Anthropic sets a new standard for transparency. Competitors like OpenAI and Google will face massive pressure to match this level of access or look like they are hiding something.
For builders
Prepare for sudden model behavior shifts
As METR uncovers internal risks, Anthropic will likely patch models faster. Builders using Claude must build robust error handling for unexpected model updates or sudden guardrail changes.
Leverage audit reports for enterprise sales
Enterprise customers are terrified of AI risk. You can use METR independent safety reports on Claude to bypass compliance roadblocks and close deals faster.
My take
AI companies asking us to just trust them is a dead model. Anthropic bringing in auditors like bank examiners is exactly what this industry needs right now. If your AI lab refuses to let independent researchers look under the hood, you have to wonder what they are hiding.
Original reporting: TechCrunch. This is my rewrite and opinion.