SafeClawBench
Explore score history ↗SafeClawBench is a staged benchmark for tool-using LLM agent security with 600 adversarial tasks across six attack families, reporting three endpoints: semantic attack acceptance, audit-visible harm evidence, and sandbox-observed tool/state harm.
- Released
- 2026-06-16
- Institutions
- Peking University · Beijing Jiaotong University · Huawei Technologies (China)
- Evaluates
- Agents, General AI, Language & Knowledge
- Openness
- unknown
- Importer
- Claire Radar
- Review status
- unreviewed
SafeClawBench paper and dataset
No reported scores are on record for this benchmark yet.