Independent Testing Enables Unfiltered Risk Disclosure
Anthropic has given the nonprofit METR employee-like access to its AI systems, allowing independent evaluators to test models for catastrophic risks without prior approval. The arrangement, based in San Francisco, fulfills a pledge to strengthen safety oversight as concerns grow over uncontrolled AI advancement. METR, which has collaborated with OpenAI, Google, and Meta, now operates with embedded staff inside Anthropic’s evaluation framework.
Breaking news
Town AI Secures $108 Million Series B, Valued at $1 Billion After $55 M Pre‑Series
Apple Prepares for Major Product Launch Event on October 28
Apple Releases Release Candidates for watchOS, tvOS and visionOS 27.2
Apple Limits iPhone Duo and iPhone 18 Pro Max to Reservation-Only Sales in Japan and South KoreaUnder the new access, METR researchers can publish findings about model dangers without needing Anthropic’s consent. This setup aims to remove barriers that might delay or suppress critical safety disclosures. The watchdog’s team works alongside internal staff but maintains authority to go public with results. Anthropic says the move supports transparency and accountability in high-stakes AI development. METR’s history of cross-industry evaluations adds credibility to the oversight process.
How Does This Access Differ From Standard Audits?
Unlike typical audits where companies control scope and timing, METR’s embedded role allows continuous, unrestricted probing of Anthropic’s systems. The access mimics internal employee privileges, including system-level insights and real-time monitoring capabilities. This depth enables detection of subtle failure modes that periodic reviews might miss. Anthropic frames the step as part of its responsible scaling policy, designed to preempt extreme risks. Critics note that true independence depends on funding and structural safeguards beyond access alone.
What specific risks is METR evaluating in Anthropic’s models? METR is assessing potential for catastrophic outcomes, such as models enabling large-scale harm or operating beyond human control, focusing on emergent behaviors in advanced systems.
Frequently Asked Questions
Can Anthropic block METR from publishing concerning findings? No, the agreement explicitly allows METR to publish results without Anthropic’s approval, ensuring unrestricted disclosure of safety concerns.
Is this access limited to current AI models or includes future versions? The arrangement covers ongoing evaluations and is expected to extend to future models as part of Anthropic’s long-term safety commitment.
