From Ars Technica · · 1m 30s
Anthropic embeds AI watchdog inside its offices
The AI safety nonprofit METR gets employee-level access to Claude's systems to assess risks, marking the first major step toward external oversight of frontier AI labs.
In brief
The AI safety nonprofit METR gets employee-level access to Claude's systems to assess risks, marking the first major step toward external oversight of frontier AI labs. Anthropic granted the nonprofit METR embedded access to its systems and hardware to evaluate AI safety risks, following through on a proposal backed…
Anthropic opens doors to safety auditors
Anthropic has granted METR, a nonprofit AI safety evaluator, employee-level access to its systems and hardware to assess risks from its Claude models.
METR embeds a team at Claude's maker
METR, which has worked with OpenAI, Google and Meta, now has six employees stationed in Anthropic's office with access to internal tools and systems.
A rare moment of unity on oversight
In September, Anthropic chief executive Dario Amodei proposed embedding external auditors in leading labs. OpenAI's Sam Altman and SpaceX's Elon Musk rallied behind the idea.
Why Anthropic cannot share everything publicly
Anthropic said stronger safeguards are needed as models grow more powerful. It cannot publicly share all safety information without benefiting malicious actors or competitors.
“As models have become more powerful, they've required stronger safeguards and more challenging risk assessment.”
The bank supervision model for AI
Anthropic compared its arrangement to bank regulation, where resident examiners work inside large banks to spot risks before they threaten the public.
“If a bank's risk-taking could threaten the public, third parties try to see that in advance.”
METR publishes findings, with limits
METR will publish its analysis without Anthropic veto, but Anthropic can require redaction of customer data, legal information and model architecture details.
The one thing
Anthropic granted the nonprofit METR embedded access to its systems and hardware to evaluate AI safety risks, following through on a proposal backed by OpenAI and SpaceX leaders.
