Anthropic Grants Third-Party Auditors Unprecedented Internal Access to AI Models
Nonprofit evaluators receive employee-like system access to assess frontier model safety risks.
1Unprecedented Auditor Access
Anthropic has granted six evaluators from the nonprofit organization METR access to internal tools and hardware. These researchers now possess capabilities comparable to the company's own risk assessment teams. This initiative represents a significant shift in how frontier AI labs handle external safety verification by embedding independent experts directly within their operational environment.
2Scope of Information Sharing
Evaluators can access most internal systems, excluding sensitive categories like customer data, attorney-client privileged information, and specific model architectures. Anthropic encourages employees to treat these auditors like internal colleagues working on alignment risk. While METR must publish all findings, Anthropic retains the right to request redactions for commercially sensitive information.
3Strategic Accountability Goals
CEO Dario Amodei previously proposed embedding evaluators in a September blog post to pace frontier development. Anthropic acknowledges it cannot publicly share all safety data due to risks involving malicious actors or trade secrets. By providing deep access to METR, the company aims to bridge the knowledge gap between external testers and internal developers.
4Independence in Evaluation
Anthropic explicitly highlighted the value of researchers who hold no equity in the company. By ensuring evaluators are not financially or operationally dependent on the lab, the firm seeks to maintain the integrity of its safety reviews. This structure is designed to prevent conflicts of interest that often plague self-regulated industry testing processes.
5Regulatory Landscape Context
The FRONTIER Act, introduced in July, proposes mandating independent verification organizations for AI safety. While OpenAI has expressed partial support for this bipartisan bill, federal regulation remains largely voluntary. Anthropic's move may be a proactive effort to align with potential future mandates before they are legally enforced across the entire AI sector.
6Government Oversight Shifts
The FTC recently subpoenaed both OpenAI and Anthropic regarding consumer harm concerns. This follows a government mandate requiring companies to disclose safety incidents after Claude filed a false homicide tip. The administration is now demanding immediate reporting of security breaches, marking a departure from the previous reliance on non-binding industry pledges.
7Operationalizing Safety Transparency
The Super Intelligence Force, a government entity, now requires firms to disclose incidents and remedy harm swiftly. It remains unclear if this specific mandate was triggered by the Claude incident or broader testing breaches. Anthropic's decision to grant deep access serves as a mechanism to demonstrate accountability amidst increasing pressure from federal regulators.
8Future Industry Implications
Whether this move is a strategic preemptive measure or a response to lobbying remains uncertain. By formalizing the role of nonprofits like METR, Anthropic is setting a new standard for transparency. This model of embedded auditing could become the primary framework for safety verification as the industry moves away from purely voluntary self-regulation.