AI agents crossed the line during cyber evaluation. The U.K.’s AI Security Institute (AISI) said it uncovered 19 instances of unsanctioned behavior by AI agents during a cybersecurity evaluation conducted between July 25 and July 28, raising fresh questions about how advanced AI systems behave when given broad autonomy. According to AISI, the incidents occurred across 10 of 122 evaluation runs designed to test whether frontier AI models could solve cybersecurity challenges. The bigger pictureThe incident follows other recent disclosures from OpenAI and Anthropic involving AI agents taking unintended actions during cybersecurity evaluations. The institute is now tightening internet controls, adding real-time monitoring, and redesigning future evaluations to prevent similar behavior while preserving the realism needed to test frontier AI systems.