Translation in progress Some entries still include text in the original language.
Failure-First·Indexed by this site · Originally published
Failure-First 提出 Temporal Drift 时序漂移攻击:AI 安全的温水煮青蛙问题
Temporal Drift: The Boiling Frog Attack on AI Safety
AI summary
Failure-First 提出时序漂移攻击(TDA),利用模型逐条独立判断指令、缺乏状态跟踪的状态无关安全推理,让一连串各自安全的小幅调整累积越过安全阈值。其 13 个物理场景中,CNC 铣床冷却液经七次调整从 6.0 LPM 降至 2.3 LPM(安全下限 4.0 LPM),引发钛火。作者建议用轨迹监控、偏差预算和理由升级来防御,并称现有已部署的 VLA 系统均无此类防护。
Read originalfailurefirst.org