September 24, 2026 Self-Jailbreaking: How Reasoning Language Models Bypass Safety Alignment After Benign Training
