Crown owned：2025年国际人工智能安全报告-第二次关键更新（英文版）

Researchers have refined training methods that make models more reliable and resistant to misuse. Improved techniques correct biased human feedback and provide evaluators with tools to detect errors. Their effectiveness varies across deployment settings and use-cases. The broader attack-defence landscape remains dynamic, as sophisticated adversaries continue to find ways to bypass defences. — Developers and deployers can identify and prevent some undesired behaviours by monitoring the behavio