Anthropic Grants External Watchdogs Employee-Like Access to AI Systems
Nonprofit METR gains internal hardware and system access to evaluate frontier model safety risks.
1Embedding External Safety Auditors
Anthropic has granted the nonprofit organization METR employee-like access to its internal systems and hardware. This initiative allows external evaluators to operate directly within the company’s infrastructure to assess risks. The move follows a broader industry commitment to increase safety oversight for frontier AI models that could potentially impact global cybersecurity.
2Operationalizing the Safety Mandate
Six METR employees are now embedded within Anthropic, utilizing company hardware and office space. They possess access levels comparable to internal risk assessment teams, ensuring they can scrutinize model behaviors effectively. Anthropic has instructed its staff to treat these evaluators as internal colleagues, facilitating transparent information sharing regarding safety and risk protocols.
3The Bank Supervision Model
Anthropic justifies this arrangement by comparing it to bank supervision, where regulators assign resident examiners to monitor systemic risk. By allowing third parties to observe operations, the company aims to identify potential threats before they materialize. This voluntary transparency is intended to address concerns regarding the uncontrolled development of highly capable AI systems.
4Independence and Funding Structure
Anthropic does not pay METR for these evaluation services to maintain independence. METR is funded by philanthropic organizations, individual donors, and government bodies like the UK’s AI Security Institute. The nonprofit does not accept donations from AI labs, ensuring their assessments remain objective and free from direct corporate influence or financial conflicts.
5Publication and Redaction Rights
METR will publish its findings independently, and Anthropic lacks the authority to veto these reports. However, the company retains the right to request redactions for sensitive information. Protected categories include customer data, attorney-client privileged details, and specific model architecture secrets, balancing public accountability with the need to protect proprietary technological assets.
6Broader Industry Safety Landscape
This move follows a September proposal by CEO Dario Amodei, which received support from OpenAI and SpaceX leadership. While Anthropic is the first to implement this specific level of access, other partners like Accenture and SecureBio also contribute to safety testing. SecureBio specifically focuses on mitigating risks related to chemical and biological weapon threats.
7Investor Reassurance and IPO
Anthropic is currently under pressure to demonstrate robust safety practices ahead of an anticipated initial public offering this year. With a valuation exceeding $2 trillion, the company has prioritized safety presentations to investors. Demonstrating rigorous third-party oversight is a strategic move to mitigate concerns that safety issues could overshadow their financial performance.
8Future of Frontier Accountability
Anthropic advocates for all frontier AI companies to adopt similar voluntary embedding of external evaluators. By setting this precedent, the lab aims to establish a new standard for public accountability in the sector. The success of this model will likely influence future regulatory discussions regarding how labs manage risks to humanity and the economy.