ClawDojo: Dynamic, Extensible Evaluation Framework for Attacks and Defenses on OpenClaw
ClawDojo frames OpenClaw-style agent security around architectural attack surface (e.g., persistent memory and third-party skills) and uses a benchmark harness to test prompt-injection and supply-chain skill abuse against candidate defenses in a controlled way.
#PromptInjection #LLMSecurity #AISecurity #Research
https://openreview.net/forum?id=MmytvTGycS
Post #2340
159
Forwarded from AISecHub
openreview.net ClawDojo: A Dynamic and Extensible Framework for Evaluating Attacks... Autonomous AI agents are evolving into continuous, local processes with operating-system level permissions. While platforms like OpenClaw exemplify this shift, granting large language models (LLMs)...