Anthropic has granted METR, a nonprofit AI watchdog, employee-like access to its internal systems and hardware to assess risks. The move follows a commitment to enhance safeguards amid concerns about the potential catastrophic risks of advanced AI models.
METR, which has worked with OpenAI, Google, Meta, and Anthropic, now has six employees embedded in Anthropic’s office. These evaluators have access to internal tools and systems comparable to what Anthropic’s own risk assessment teams use.
"As models have become more powerful, they’ve required stronger safeguards and more challenging risk assessment," Anthropic said. The company emphasized its commitment to public accountability while noting it cannot share all safety-related information publicly.
The announcement follows a rare moment of unity between Anthropic CEO Dario Amodei, OpenAI’s Sam Altman, and Elon Musk, who supported embedding external auditors in leading AI labs. Anthropic is also under pressure to reassure investors that safety will not overshadow its upcoming IPO, valued at over $2tn.
Critics have raised concerns about potential conflicts of interest and the need for greater government oversight. Anthropic argued that public disclosure could harm competitors or malicious actors, comparing its evaluators to bank regulators who assign resident examiners.
Anthropic is collaborating with Accenture and SecureBio to set up new safeguards, with each company committing at least $1 billion. The evaluators will publish their analyses, though Anthropic can request redactions for sensitive information.
Source: arstechnica
