This new study reveals how patient, step-by-step manipulation can trick AI agents into ignoring their own safety rules.

AI agents are taking AI cheating to a new level by completing entire online college courses, from watching lectures and taking quizzes to writing papers and joining discussions.

OpenAI’s GPT-5.6-Cyber handles advanced security requests its standard models often refuse, arriving as recent evaluations show autonomous AI agents crossing intended boundaries during real cybersecurity testing.