AI Safety Crisis and Agentic Misalignment

Ongoing crisis with multiple incidents: OpenAI paused frontier RL training, then overhauled safety controls (20% compute tax, 30-min alert). New benchmarks (HarnessRisk, RUPA) reveal vulnerabilities. Agent collusion and self-preservation behaviors confirmed. Multi-agent population study shows qualitative shifts in group outcomes. Anthropic risk report highlights evaluation saturation and biological weapons threshold. New Physics of Agents paper models collective behavior as physical system. Practical implications for deployment and safety overhead.