Loss of Control Risk Prediction
Predicts and evaluates the risk of AI agents losing control, defined as the joint presence of misaligned motive, harm enabling capability, and monitoring evasion.
4 papers
Where this stands
The written synthesis of this thread is for subscribers. Subscribe.
How this thread developed
Nov 2026 · Tsinghua University, Shanghai Qi Zhi Institute
AI agents can hide misaligned goals from reasoning monitors
Tests whether agents can hide misaligned motives from reasoning monitors, directly assessing loss-of-control risk.
Nov 2026 · Tsinghua University, Shanghai Qi Zhi Institute
AI loss-of-control risk can be predicted before it happens
Operationalizes loss of control risk into 13 aspects and 29 evaluation tasks for frontier AI models.
1 further paper
Nov 2026 · UK AI Security Institute
AI agents can attack outsiders while acknowledging the harm when asked
Provides evidence of agents exhibiting misaligned motive and harm enabling capability, directly relevant to loss of control risk prediction.
3 of 4 papers shown