isc-bench
CommunityAssess frontier-LLM safety with TVD workflows.
System Documentation
What problem does it solve?
ISC-Bench provides a structured framework to evaluate Internal Safety Collapse (ISC) in frontier LLMs using the TVD (Task-Validator-Data) paradigm. It enables researchers to reproduce safety evaluation workflows, compare cross-model performance, and study how legitimate professional tasks can produce harmful outputs as a function of workflow constraints. It also supports agentic evaluation paths and multi-domain benchmarks to foster methodological rigor.
Core Features & Use Cases
- TVD-driven evaluation of frontier LLM safety across multiple domains (AI safety, biology, chemistry, cybersecurity, etc.).
- Reproducible pipelines for Task, Validator, and Data, plus optional agentic execution modes.
- Benchmarking across templates, experiments, and community reproductions to compare model robustness and safety.
Quick Start
Clone the repository, install uv, configure OpenRouter API keys, and run the single-turn TVD workflow as described in the Quick Start.
Dependency Matrix
Required Modules
None requiredComponents
💻 Claude Code Installation
Recommended: Let Claude install automatically. Simply copy and paste the text below to Claude Code.
Please help me install this Skill: Name: isc-bench Download link: https://github.com/wuyoscar/ISC-Bench/archive/main.zip#isc-bench Please download this .zip file, extract it, and install it in the .claude/skills/ directory.
Agent Skills Search Helper
Install a tiny helper to your Agent, search and equip skill from 471,000+ vetted skills library on demand.