AutoRedClaw combines a strict quality-checking protocol with execution-based validation to ensure correctness, realism, and effectiveness of generated samples.
AutoRedClaw automates red-teaming for persistent claw-style agents
The framework generates end-to-end attacks including benign tasks, adversarial instructions, and executable environments to evaluate threats to agent configuration, memory, and skills.
Big Tech
Hao Li · Tongyan Hu · Weidi Luo · Shanghao Shi · Xiaogeng Liu · Aiden Suh · +5 more
Washington University in St. Louis · Johns Hopkins University · National University of Singapore · University of Georgia · Independent Researcher
Research Digest··2 min read
Li et al.
§