Anthropic is giving AI safety evaluators “employee-like” access to its systems.
Six employees at the AI safety organization METR will gain access to hardware and internal tools “roughly on par” with what Anthropic’s own risk assessment team sees. While Anthropic says METR will publish reports about Anthropic’s risk levels, incidents, and practices, it can still instruct the safety evaluator to redact certain information, such as the inner workings of its models or customer data.
TL;DR
- Six METR employees will have access to Anthropic's hardware and internal tools.
- This access level is comparable to Anthropic's own risk assessment team.
- METR is intended to publish reports on Anthropic's risk levels, incidents, and practices.
- Anthropic can instruct METR to redact specific information, including model inner workings and customer data.