wuyoscar/ISC-Bench

Provides both single-turn templates and autonomous agent-based execution modes for systematically triggering safety failures in frontier LLMs through incomplete professional workflows (e.g., biomedical analysis, chemical synthesis) rather than adversarial prompts. The benchmark uses domain-specific task templates across biology, chemistry, and epidemiology with layered evaluation scaffolds, paired with an ISC Arena leaderboard tracking vulnerability patterns across models via API-based reproducibility and community submissions through GitHub Issues.

677 stars. Actively maintained with 289 commits in the last 30 days.