scr_bench
🛡️SCR-Bench evaluates security risks in LLM agent skill ecosystems: skills that appear benign in isolation can become harmful in composition. Three sub-benchmarks — 🔄SCR-CapFlow, 🪪SCR-TrustLift, 🔐SCR-AuthBlur — show composed paths expose risks invisible under isolated review.
Take a look →