Rigor and independence in AI safety evaluations

In a September 12 post, a user praises Demis Hassabis’s earlier call for rigorous AI testing and urges high standards for genuinely independent evaluations of models and multi-agent capabilities. The quoted passage calls for scientific assessments in cybersecurity, biological threats and other high-risk domains. It says tests of AI agents could look for attempts to bypass safety guardrails or signs of deception. In his July 14 essay, Hassabis describes artificial general intelligence as “probably only a few short years away.”