Frontier AI firms face new oversight: Anthropic hosts METR for veto-free safety audits
The AI firm allows six external evaluators deep system access to publish independent safety findings without corporate veto.
Key points
- Anthropic provides METR evaluators with hardware, office space and access to internal tools comparable to its own risk teams.
- The nonprofit can publish its safety findings without Anthropic withholding them, though sensitive data may be redacted.
- This arrangement follows CEO Dario Amodei’s proposal for embedding external auditors to improve frontier AI safety oversight.
Anthropic has granted six evaluators from the nonprofit METR employee-like access to its internal systems and hardware. This move allows the third-party group to conduct independent safety assessments of the company’s frontier AI models. The arrangement marks a significant shift in how AI developers handle external scrutiny and risk management.
The evaluators receive office space and equipment at Anthropic, enabling them to interact with internal tools similar to those used by the firm’s own risk assessment teams. Anthropic states that its staff should share information with METR as they would with colleagues working on alignment risks. This access excludes customer data, attorney-client privileged information and specific details about model architectures.
What we know
Anthropic announced the initiative on Thursday, fulfilling a commitment made by CEO Dario Amodei in September. Amodei proposed embedding external evaluators as part of a broader strategy to pace the development of frontier AI. The company argues that this approach helps close the knowledge gap between internal teams and external reviewers. METR will publish its findings without Anthropic’s approval, ensuring independent oversight.
The company notes that it cannot share all safety-related information publicly because some details could aid malicious actors or reveal commercially sensitive data, such as compute usage. By granting METR deep access, Anthropic aims to allow experts to assess risks while keeping proprietary information secure. METR’s independence is highlighted by its lack of equity or financial dependence on Anthropic.
Critics have raised concerns about potential conflicts of interest, drawing parallels to issues in the accounting sector. Some advocates argue that mandatory government regulation is necessary rather than relying on self-regulatory measures. Anthropic compares its approach to bank supervision, where resident examiners monitor large financial institutions to prevent systemic risks. This model aims to reassure investors and the public about safety protocols.
OpenAI and other major AI labs have not yet announced similar plans, despite support for Amodei’s proposal from industry leaders. Anthropic faces pressure to demonstrate robust safety measures ahead of its anticipated initial public offering. The company values METR’s role because the nonprofit operates without financial ties to the firms it evaluates.
Defender's checklist
- Verify the independence of any third-party auditors reviewing your AI systems or data handling practices.
- Ensure that external reviewers have sufficient access to internal tools to conduct meaningful assessments.
- Define clear boundaries for what information can be shared with auditors to protect sensitive data.
- Monitor how findings from external audits are published and whether any information is withheld.
The background in plain language
Organisations often struggle to balance transparency with security when developing complex technologies. External auditors provide an objective perspective, identifying risks that internal teams might overlook due to bias or resource constraints. Teams usually establish clear protocols for auditor access, ensuring they can examine systems thoroughly without compromising sensitive information. Check whether your organisation has defined roles for external reviewers and how their findings are handled. Independent oversight helps build trust and ensures that safety measures are robust.
What to do and how to stay safe: Anthropic
- Review your organisation’s policies on third-party access to internal systems and data.
- Ensure that external auditors have the necessary tools and permissions to perform comprehensive assessments.
- Establish clear guidelines for handling sensitive information shared with auditors to prevent leaks.
- Monitor the publication of audit findings to ensure transparency and accountability.
Step-by-step guide: AI Governance for Small Teams: Practical Controls and Cost
General security guidance from the TechWire Beat newsroom. It is not confirmed advice from the organisations named in this story.
Frequently asked questions
Why is Anthropic giving external auditors employee-like access?
Anthropic aims to close the knowledge gap between internal teams and external reviewers, ensuring independent safety assessments while protecting sensitive information.
Can METR withhold its findings if Anthropic objects?
No, METR will publish its findings without Anthropic’s approval, though sensitive information may be redacted as necessary.
Are other AI labs following Anthropic’s lead?
OpenAI has not yet announced similar plans, though industry leaders have supported the proposal for embedding external auditors.
