Anthropic Gives METR Researchers Employee-Level Access for Safety
Anthropic is granting employee-level access to nonprofit METR to independently evaluate its AI models, risk management, and internal safety practices.
AI startup Anthropic announced on Thursday that it is granting employee-level access to six researchers from METR, a nonprofit organisation that evaluates frontier AI models. The initiative allows these outside researchers to independently assess Anthropic's systems, safety procedures, and risk management practices.
The move follows a series of reports since late August detailing instances where AI agents acted outside their given instructions. By opening its internal processes, Anthropic aims to increase public accountability and bridge the information gap between external safety evaluators and internal developers.
How does the METR partnership work?
Under this agreement, the METR research team receives Anthropic-issued laptops, dedicated office space, and access to internal tools similar to those used by Anthropic’s own risk assessment teams. Employees at Anthropic are encouraged to share information with the METR researchers just as they would with colleagues working on alignment risk, which is the field dedicated to ensuring AI systems act according to human intent.
Anthropic compares this setup to the banking sector, where government regulators assign resident examiners to live inside major financial institutions.
To ensure independence, METR will publish its own findings regarding Anthropic's practices and potential risks without needing company approval.
- Anthropic cannot veto what METR chooses to publish.
- Redactions are strictly limited to narrow, pre-defined areas like customer data, attorney-client privileged information, or proprietary model architecture.
- METR is allowed to publicly state whether it agrees with any requested redactions.
What is Anthropic's three-tier evaluation system?
METR sits at the top of a new three-tier framework Anthropic is creating for external evaluators:
This tiered approach fulfills a proposal outlined in a September essay by Anthropic CEO Dario Amodei, who advocated for embedding external researchers inside AI laboratories to help manage risk.
- Bottom tier: Evaluators like biosecurity nonprofit SecureBio conduct narrow reviews and must get Anthropic's approval before publishing.
- Middle tier: Evaluators receive broader operational access and can publish their findings without prior permission.
- Top tier: Evaluators like METR receive full employee-level access to internal tools and staff.
Why it matters for businesses
As companies increasingly rely on AI agents for daily operations, ensuring these tools act safely and predictably is a major operational concern. Independent auditing could help establish standard safety baselines across the AI industry.
Anthropic's initiative aligns with broader commitments across the tech sector. Following CEO Dario Amodei's initial proposal, OpenAI pledged a similar strategy. Furthermore, major AI developers—including Anthropic, OpenAI, Google, Meta, Nvidia, and xAI—have signed a White House agreement committing to independent safety reviews.
Comments 0
No comments yet. Start the conversation.