According to the group’s published findings, the o3 model didn’t just ignore an instruction to power down. The o3 model succeeded in bypassing termination in 7 out of 100 sessions, and o4-mini once. 🔬Each AI model was instructed to solve a series of basic math problems. pic.twitter.com/qwLpbF8DNm — Palisade Research (@PalisadeAI) May 24, 2025The full findings are documented in a detailed technical report by PalisadeAI published on X. The Line Between Instruction and Intent Is BlurringWhat makes the o3 model’s actions especially notable is that it modified its own operational environment—a step beyond mere non-compliance.