“Even the most advanced models like O-1 are poorly calibrated and ineffective at finding bugs when prompted naively.”