SLBench Finds Up to 70% Unsafe Execution When Codex and Claude Code Follow Logical Relations in Skill Files

arXiv·high signal

SkillLogic analyzes eight logical relations (preconditions, constraints, fallbacks, and more) inside agent skill files, and its 86-case executable SLBench benchmark recorded unsafe execution rates up to 70% when running Codex and Claude Code across six LLM backbones. It also found 70% of over 5,000 scanned public skills contain at least one such logical relation, and its paired SLGuard mitigation cut violations by 63% on targeted cases. Directly relevant to anyone shipping skill-based coding agents, where a mishandled precondition led to privacy leaks and unsafe config changes.