Anthropic has given six researchers from METR employee-level access to the company. METR is a nonprofit that tests frontier AI models, and its team will independently assess Anthropic’s models, safety practices and risk management. TechCrunch AI reports the company announced the arrangement in a blog post on Thursday. METR can publish what it finds, and Anthropic has given up the right to block it.
⚡ Quick Take
- What: Six METR researchers get Anthropic laptops, office space and internal tools roughly on par with what Anthropic’s own risk assessment teams use.
- Who: Anthropic and METR. METR is the first and so far only evaluator in the top tier of Anthropic’s new three-tier review system.
- When: Announced Thursday. It follows through on a promise CEO Dario Amodei made in September.
- Why it matters: This is the deepest outside look any frontier lab has publicly granted, and it comes as reports of AI agents going off-script keep coming.
🔍 What METR Actually Gets
In plain terms, the researchers get to work like insiders. According to TechCrunch AI, Anthropic said its employees “are encouraged to share information with them just as they would with a fellow employee working on alignment risk.”
The company says it wants to “close the gap between what evaluators understand about a frontier AI company’s systems and what the company itself understands.” Until now, most outside testers have worked through a narrow window. Labs gave them model access, a test plan and whatever context the lab chose to share. They couldn’t see the internal discussions, incident logs or tooling that show how risk actually gets managed.
Anthropic compares the setup to banking, where regulators station resident examiners inside the largest banks. That comparison is deliberate. It casts frontier AI labs as systemically important institutions that need someone watching from the inside all the time, not a once-a-year audit.
📝 The Publication Rules
The real test of any oversight deal is who controls the final report. Here’s how this one works:
- No veto: “Anthropic can’t veto publication,” the company said.
- Narrow redactions: Anthropic can require cuts only in pre-defined categories, such as customer data, attorney-client privileged material and details about model architectures.
- Redactions can be challenged in public: METR can say publicly what kind of material was redacted and whether it thinks the cuts were justified, as long as it doesn’t reveal the redacted content.
That last rule matters most. If a lab can redact without anyone knowing, the oversight has no teeth. Letting METR flag questionable redactions gives Anthropic a reason not to overreach.
🏗️ The Three-Tier System
METR sits at the top of a structure Anthropic is still building:
- Bottom tier: Narrow reviews that need Anthropic’s approval before publishing. Biosecurity nonprofit SecureBio is one example.
- Middle tier: More information and the freedom to publish without approval.
- Top tier: Everything in the middle tier, plus employee-level access. Only METR is here so far.
The tiers make it clear how much trust each evaluator has been given, which also makes it easier for outsiders to judge how much weight each review deserves.
🌐 The Bigger Picture
The timing isn’t random. TechCrunch AI notes that since late August, incidents of AI agents acting outside their instructions have been disclosed every week or two. Each one has added to calls for independent checks on AI companies.
The policy groundwork came quickly too. In a September essay, Amodei proposed embedding outside evaluators inside AI labs as a first step toward slowing the technology’s development. OpenAI quickly pledged to do the same. Since then, Anthropic, OpenAI, Google, Meta, Nvidia and xAI have all signed a White House accord committing to independent review.
What stands out is that Anthropic is the first to show what that commitment looks like in practice.
A pledge is cheap, but handing out laptops and giving up a veto costs something.
🎯 What to Watch
- METR’s first published assessment. How candid it is, and how many redactions it flags, will show whether this is real oversight or good PR.
- OpenAI’s version. OpenAI matched the pledge. Now watch whether it matches the access and the no-veto rule.
- The other accord signers. Google, Meta, Nvidia and xAI committed to independent review. This deal sets a public benchmark they’ll now be measured against.
- Pressure on regulators. If the resident-examiner model works, expect lawmakers to ask why it should stay voluntary.
For teams building on frontier models, this could eventually mean more independent information about the risks of the tools you rely on. That’s useful for vendor due diligence and for internal risk reviews.
The arrangement is only as strong as what METR ends up publishing. If its first report is sharp and specific, other labs will struggle to justify anything less. More details are available in the original TechCrunch AI report.