SeerGuard: A Safety Framework for Mobile GUI Agents via World Model Prediction
SeerGuard tries to catch dangerous mobile-agent actions before they run by predicting their likely GUI consequences.
The paper frames current safeguards as too reactive for phone-control agents, where one bad tap can cause irreversible harm. SeerGuard adds instruction-level screening plus action-level risk checks against the current GUI state. Its safety-augmented world model combines next-state prediction with risk assessment. In experiments, the authors report better safety-utility and lower risk-cost scores on Qwen3-VL-8B-Instruct. HF Daily Papers' note
The paper frames current safeguards as too reactive for phone-control agents, where one bad tap can cause irreversible harm. SeerGuard adds instruction-level screening plus action-level risk checks against the current GUI state. Its safety-augmented world model combines next-state prediction with risk assessment. In experiments, the authors report better safety-utility and lower risk-cost scores on Qwen3-VL-8B-Instruct. HF Daily Papers' note
score 5