Who tests the testers? Systematic enumeration and coverage auditing of LLM agent tool-call safety

The researchers propose SafeAudit, a meta-audit framework that audits the coverage gaps of existing safety benchmarks by systematically enumerating the tool-call flows of LLM agents. The framework uses an LLM-based enumerator to generate valid tool-call workflows and diverse user scenarios, and introduces rule-resistance, a non-semantic quantitative metric, to distill safety rules from existing benchmarks and identify unsafe interaction patterns not covered by the rules. Across 3 benchmarks and 12 environments, SafeAudit finds that over 20% of residual unsafe behaviors are not exposed by existing benchmarks, and coverage grows monotonically with increasing test budget.