ConfGuard: a physics-grounded reference monitor for predictive-maintenance large-language-model agents
- Published
- Record updated
Summary
ConfGuard is a reference monitor for LLM agents that run predictive-maintenance loops, where injected retrieved text or bounded sensor perturbations can corrupt maintenance decisions. Its conformal monitor checks each remaining-useful-life prediction against degradation physics and a physics-policy validator checks the agent's urgency decision, reducing a covert attacker's reach on healthy assets from 0.529 to 0.046 on 10,196 C-MAPSS test windows. An adaptive attacker still defeats a Mahalanobis detector on 0.504 of windows and adversarial training on 0.485, and the authors report an attack class the projection misses.
Mitigation
ConfGuard itself, a reference monitor with a physics-grounded policy: a conformal monitor testing each prediction against degradation physics (capped remaining life never rises and falls at most one cycle per cycle) and a physics-policy validator checking the agent's maintenance-urgency decision.
Related items
- HighCVE-2026-75612: yt-mcp arbitrary file read through add_attachment file_pathSimilar attack · NVD/CVE Database
- HighCVE-2026-130121: The AI Puffer - MCP Server for Claude & ChatGPT, with AI Chatbot plugin for WordPress is vulnerable to Stored…Similar attack · NVD/CVE Database
- InfoExternal sandboxes for coding agentsSimilar attack · Simon Willison's Weblog
- InfoEvolution Under Fire: Evolving Agent Safety Harnesses with Adaptive Injection StressorsSimilar attack · Arxiv (cs.CR + cs.CL + cs.LG)
- MediumCVE-2026-130358: AdminNeo before 5.9.1 contains a missing authentication vulnerability in the ClaudeSqlPlugin, GeminiSqlPlugin, and…Similar attack · NVD/CVE Database