METR Researchers Gain Employee-Level Access to Assess Unreleased Models and Publish Independent Risk Findings. Anthropic, one of the labs that pushed for embedded evaluators, said in a blog post that it is giving the independent evaluator Metr employee-like access to its models before deployment. This level of access is close to what Metr said are best practices for its assessors to understand how highly capable models behave.
First seen on govinfosecurity.com
Jump to article: www.govinfosecurity.com/anthropic-gives-ai-safety-evaluators-insider-access-a-33134
