Anthropic Grants External AI Watchdog Employee-Level Access to Enhance Safety Oversight

Image: Ars Technica · Source

Anthropic has embedded an external watchdog team within its offices, giving them extensive access to assess AI model risks independently, marking a major step in AI safety transparency and accountability.

Anthropic, a leading artificial intelligence developer, has taken a significant step to bolster safety oversight by granting an external watchdog organization what it calls "employee-like access" to its internal systems and hardware. This initiative aims to transparently evaluate the risks associated with its AI models, including Claude, by embedding the watchdog team directly within Anthropic's offices. The external evaluators belong to METR, a nonprofit with prior experience working alongside major AI labs such as OpenAI, Google, and Meta. By having full access comparable to Anthropic’s own risk assessment teams, METR staff can independently audit, review, and analyze safety issues.

This move follows broad industry and public concern over the escalating potential dangers posed by increasingly powerful AI systems, ranging from economic disruption and cyber threats to existential risks. Anthropic CEO Dario Amodei's push for third-party evaluations had previously drawn support from influential AI figures including OpenAI's Sam Altman and SpaceX's Elon Musk, highlighting a rare consensus on the importance of robust external oversight.

Anthropic has emphasized that while it seeks public accountability in its safety work, direct public disclosure of all sensitive information could expose vulnerabilities to malicious actors or competitors. Instead, it has opted for a model analogous to banking supervision, where regulators station examiners within large banks to assess risks first-hand and proactively. METR evaluators have the authority to publish their findings without needing Anthropic’s approval, though the company can request redactions for sensitive data categories such as customer information, legal communications, or proprietary model details.

To reinforce this oversight, Anthropic has equipped METR with company hardware, office space, and unrestricted access similar to that of internal safety teams. Employees at Anthropic are encouraged to interact with METR reviewers as they would with colleagues on risk issues, fostering collaborative transparency. Notably, this arrangement is undertaken without financial compensation from Anthropic to METR, which operates solely on philanthropic and government funding, preserving its independence.

This development arrives as Anthropic prepares for a high-profile initial public offering, estimated at over $2 trillion in valuation this year. The company is keen to demonstrate that AI safety and rigorous risk management remain central priorities despite commercial pressures. Beyond METR, Anthropic continues collaborations with other safety partners like Accenture, which is involved in extensive safety testing investments, and SecureBio, which assesses risks related to chemical and biological threats.

Anthropic has called on other frontier AI firms to adopt similar embedded evaluator programs, emphasizing voluntary transparency to collectively enhance industry safety standards. METR expressed readiness to extend its oversight efforts across the AI landscape to provide improved public evidence and assurance regarding AI safety.

In summary, Anthropic's unprecedented granting of extensive internal access to an external watchdog marks a notable advancement in AI governance. It sets a potential precedent for embedding independent evaluators within AI companies to proactively identify and mitigate risks amid growing calls for stronger regulatory frameworks on advanced AI technologies.