"Elementary, My Dear Watson." Detecting Malicious Skills via Neuro-Symbolic Reasoning across Heterogeneous Artifacts
The paper 'Elementary, My Dear Watson' introduces MalSkills, a robust solution for detecting malicious AI skills. It addresses threats that are spread across various artifacts like prompts and code scripts by using a holistic approach. The method maps data flow to identify complex dangers such as credential theft and data exfiltration, providing a foundation for securing the agentic supply chain.