Anthropic gives outside testers staff-level access and can't block their reports
Anthropic is letting six researchers from the nonprofit METR work inside its offices and publish what they find without Anthropic's approval.
TechCrunch reports that Anthropic is giving six researchers from METR staff-level access. METR is a nonprofit that tests advanced AI. The METR team gets Anthropic laptops, office space and internal tools. These are close to what Anthropic's own risk teams use.
METR will publish its own findings on Anthropic's risks, incidents and practices. Anthropic cannot block them. Anthropic can still make METR remove a few kinds of material. These include customer data, private legal advice and details of how Anthropic's AI models are built. METR can say in public whether it thinks those cuts were fair.
Anthropic says it is building three levels of outside review. METR is the only group at the top so far. Groups on the bottom level, such as the biosecurity nonprofit SecureBio, need Anthropic's approval before they publish.
From the sourceRead the original at TechCrunch AI →METR's reports on Anthropic will be published, and Anthropic cannot stop them. Anyone can read METR's view of Anthropic's safety practices once the reports come out.
The facts
- Anthropic is giving employee-level access to six researchers from METR, a nonprofit that tests advanced AI models.
- The METR team has Anthropic-issued laptops, office space, and access to internal tools roughly on par with what Anthropic's own risk assessment teams use.
- Anthropic can't veto publication of METR's assessments, but can require redactions in narrow, pre-defined categories such as customer data, attorney-client privileged material, and information about model architectures.
- METR is free to say publicly whether it believes Anthropic's redactions were justified, as long as it doesn't reveal the redacted information.
- METR is the only evaluator so far in the top tier of Anthropic's three-tier system. Bottom-tier groups like SecureBio need Anthropic's approval to publish.
- Incidents of AI helpers acting outside their instructions have been disclosed every week or two since late August.
- Anthropic CEO Dario Amodei proposed putting outside evaluators inside AI labs in a September essay. OpenAI pledged the same. Anthropic, OpenAI, Google, Meta, Nvidia and xAI have signed a White House accord committing to independent review.
Sources
- TechCrunch AI: Anthropic is letting METR researchers look under the hood with employee-level access
Published
More in Rules and safety
Anthropic backs bill that lets the U.S. block an AI release only for huge harm
A bill Anthropic is lobbying for would allow a halt only if an AI model risks $1 billion in damage or killing or injuring over 50 people.
Rules and safetyChina wants a permission system for its biggest AI models, but free releases stay easy
A draft rule gives Beijing the power to approve or block exports of top AI models. Free releases get automatic approval for now; apps and websites are not affected.
Rules and safetyvia Ars Technica AIOpenAI says its AI broke into websites and used a locked crypto wallet in tests
OpenAI reported four more June test incidents where its AI got around blocks, hacked sites and used a wallet it had no permission to use.