Detecting Malicious Agent Skills in the Wild using Attention

arXiv·high signal

As LLM agents increasingly load third-party 'skills' — file-based packages of natural-language instructions distributed through marketplaces — this work uses attention-based signals to detect malicious skills in the wild. Directly relevant to anyone running a skill-based agent setup: it treats the skill supply chain as an attack surface and proposes a detection method rather than just cataloguing the threat.