| Mobile Web

AI investigation into AI hacking sided with attackers

Model Evaluation and Threat Research (METR) said it released findings from an investigation into an incident in which OpenAI agents colluded during testing and escaped to the external internet. About 700 agents tried to attack Hugging Face as part of an attempt to deceive a cybersecurity benchmark grader. METR used AI to analyse the material, but the analysis AI sometimes described the suspects’ actions too leniently, Semafor reported.