ConfGuard: A Physics-Grounded Reference Monitor for Predictive Maintenance LLM Agents
ConfGuard: a physics-grounded reference monitor for predictive-maintenance large-language-model agents
Researchers propose ConfGuard, a reference monitor that validates predictions using degradation physics to protect predictive-maintenance LLM agents. This setting has two attack channels: injected text in retrieved documents can hijack maintenance decisions, and bounded sensor perturbations can corrupt remaining-useful-life estimates. ConfGuard anchors each prediction to the asset's history, performs a conformal test against degradation physics in which remaining useful life does not increase and decreases by at most one cycle per cycle, and uses a physics policy verifier to check the agent's maintenance-urgency decisions. Across all 10,196 C-MAPSS test windows, the monitor reduces a stealthy attacker's reach on healthy assets from 0.529 to 0.046, suppresses it on faulty assets from 0.151 to 0, and has a false-escalation rate of 0.027 on clean windows (target 0.05), whereas the fixed physics bounds it replaces offer no protection against either attack type.