Anthropic announced this week that six researchers from METR, a nonprofit that tests frontier models, will get employee-level access to independently assess its models, safety practices and risk management. They get company laptops, office space and internal tools at roughly the level of Anthropic's own risk teams. METR will publish its own findings, and Anthropic has given up the right to block them.
That last part is the interesting bit. Anthropic can still require redactions, but only in narrow, predefined categories: customer data, attorney-client privileged material and information about model architectures. METR can say publicly whether it thinks a redaction was justified, as long as it doesn't reveal what was removed.
Why this matters if you only consume the API
You probably don't care how a lab runs its internal reviews. You care whether the thing you ship on top of it keeps working and doesn't embarrass you. But the context is worth knowing. According to TechCrunch, disclosures of AI agents acting outside their instructions have surfaced roughly weekly since late August, and that is what is pushing labs toward independent checks. OpenAI quickly pledged something similar, and several labs have signed a White House accord committing to independent review.
The reporting also describes a three-tier system for evaluators. Narrow reviews need Anthropic's approval before publishing. A middle tier publishes without approval. The top tier adds employee-level access, and right now METR is the only one in it.
So the lab layer is getting an audit trail. Is yours?
The gap is usually on our side
Most teams I talk to can tell you which model they call. Far fewer can answer these:
- What was the last model version change that touched production, and who approved it?
- Which of our agents can write to something, and what stops them?
- If a vendor publishes a bad safety report next month, what do we do on Monday?
Lab-level oversight doesn't answer any of that. Your agent runs inside your permissions, on your data, with your prompts. An independent evaluator at the lab will never see your edge cases.
What to do with this
First, put vendor evidence into procurement. When you pick or renew a model provider, ask which independent evaluators have access, what they have published, and what the vendor is allowed to redact. A vendor that can't answer clearly is telling you something.
Second, copy the structure internally. Give someone outside the team that built the feature read access to prompts, tool permissions and run logs, and let them report without asking permission. It doesn't need to be a full audit function. One senior engineer with a monthly review slot catches a lot.
Third, keep your own evals. Public reports tell you how a model behaves in general. A small regression suite built from your real tickets tells you whether the Tuesday update broke your workflow. Run it on every model change, treat it like a dependency bump, and store the results.
The takeaway
Independent oversight is moving inside the labs, which is good news. It doesn't move the responsibility for your product. The teams that will handle the next agent incident well are the ones that already have logs, permissions and a review habit in place.
We're here to help founders and teams design and build digital products that are built to scale with you, not slow you down. If you're looking to build something, get in contact with us today!