Anthropic grants METR employee-level access to audit frontier models

A precedent in AI governance

Anthropic has announced the granting of internal permissions to six members of METR (Model Evaluation and Threat Research), an independent non-profit organization. This measure, described as 'employee-level' access, allows external evaluators to operate within the company's systems with unprecedented autonomy in the industry.

The stated objective is to close the gap between the internal understanding of AI systems and the ability of auditors to verify them. By integrating these evaluators into the daily workflow, the company seeks to ensure that conclusions regarding risks, incidents, and operational practices are based on direct observation rather than reports filtered or limited by the company itself.

This structural shift represents a fundamental departure from traditional audit methodologies, which typically rely on periodic reviews or limited data sharing. By embedding external evaluators directly into the operational fabric of the company, the initiative aims to transform safety verification from a retrospective exercise into a continuous, real-time process. This integration allows for the immediate detection of anomalies that might otherwise remain obscured within complex system interactions, thereby enhancing the overall reliability of the safety assessments conducted by the organization.

What 'employee-level access' entails

In practice, this access translates to the provision of corporate hardware, physical space in Anthropic's offices, and credentials to use internal risk assessment tools. METR researchers can interact with models and agent monitoring systems in the same manner as an internal security team technician.

However, this access is not total. Customer data, information protected by professional secrecy, and the detailed technical architecture of the models are explicitly excluded. This distinction is crucial for balancing transparency with the protection of intellectual property and user privacy.

The exclusion of detailed model architecture and customer data serves as a critical boundary that defines the limits of this transparency. While evaluators can observe the behavior and outputs of the systems, they do not have access to the underlying code or proprietary training methodologies. This separation ensures that the integrity of the intellectual property is maintained while still allowing for rigorous external scrutiny of the system's performance and potential risks, creating a balanced approach to both innovation and safety.

The three-tier auditing framework

Anthropic is designing a tiered access system for other evaluators. The first level, where SecureBio currently operates, involves research with a very limited scope and requires prior authorization to publish. The second level grants more information and allows publication without prior approval, subject to specific redactions that the auditor can publicly challenge.

METR operates at the third level, the most advanced. This level combines full technical access (similar to an employee) with publication rights analogous to the second level. The company has indicated it is in talks with other entities, such as Hugging Face and Faculty (Accenture's AI subsidiary), to integrate them into this scheme, although their current rights are more restricted.

The tiered framework is designed to accommodate varying levels of trust and operational scope among different external entities. By establishing clear distinctions between limited, expanded, and full access, the company creates a scalable model that can be adapted to the specific needs and capabilities of different auditing organizations. This structured approach not only facilitates the integration of new partners but also provides a clear pathway for existing evaluators to gain greater autonomy and influence over the publication of their findings as their relationship with the company deepens.

Independence and controversies

Although METR defines itself as independent and does not receive direct funding from the labs it evaluates, there are criticisms regarding its autonomy. The organization employs researchers who previously worked at Anthropic, OpenAI, and DeepMind, and its funding comes largely from foundations linked to the effective altruism movement, with which Anthropic CEO Dario Amodei maintains family and professional ties.

Critics such as David Sacks have pointed out that this proximity could compromise the objectivity of the audits. However, METR has emphasized that its methodology is based on the scientific evaluation of autonomous capabilities and that its work, such as the review of Anthropic's agent monitoring systems, has identified real vulnerabilities that the company has since corrected.

The debate surrounding the independence of the evaluators highlights the complex web of relationships that exist within the AI research community. While the formal separation of funding and operational control is a key pillar of the organization's credibility, the personal and professional connections between key figures in the industry and the evaluators raise questions about potential biases. Addressing these concerns requires not only transparent methodologies but also a robust mechanism for public scrutiny and accountability to ensure that the audits remain objective and unbiased.

Implications for the industry and regulation

This initiative is framed within Dario Amodei's proposal to 'pace the frontier' of AI. By establishing a verifiable standard of transparency, Anthropic seeks to influence future regulation, arguing that labs must assume the responsibility of demonstrating the safety of their models before governments impose coercive restrictions.

For the rest of the industry, this move poses a dilemma: adopt a similar level of transparency to maintain regulatory competitiveness or maintain control over technical information. METR's ability to publicly question redactions imposed by Anthropic adds a layer of accountability that could become a standard requirement for frontier models in the coming years.

The potential for this model to become an industry standard depends on its ability to demonstrate tangible improvements in safety and risk mitigation. If the findings from these embedded audits lead to significant corrections and enhancements in model behavior, it could set a new benchmark for what is expected of frontier AI developers. This shift would likely influence regulatory discussions, providing a concrete example of how voluntary transparency can complement and potentially preempt more stringent government mandates, thereby shaping the future landscape of AI governance.

Keep exploring

Corrections and transparency

If you spot incorrect or outdated information, send us a correction with the page and source we should review.