A certifying, model-free, minimal-intervention runtime shield for LLM agents: it repairs an unsafe tool-call trajectory with the least-cost enactable edit instead of blocking, and certifies every intervention for re-verification without the model.
pythonsecurityltltemporal-logicformal-methodsai-safetyinterpretabilityruntime-verificationmodel-freetamper-evidenttsetlin-machineguardrailsllmruntime-enforcementllm-agentsagent-safetyminimal-interventionshield-synthesisedit-automatacertifying-algorithms
-
Updated
Jun 28, 2026 - Python