This paper introduces HealBench and HealGuard, a benchmark and guardrail for trustworthy runtime error healing in real-world repositories. Researchers evaluate LLM-based healing code and detect potential safety issues with existing agents.