Study: How AI Agent "Skills" Leak Your Credentials

A 2026 empirical study by Chen et al. reveals that third-party "skills" integrated into AI agents frequently leak sensitive credentials. Analyzing over 17,000 agent skills, researchers identified 1,708 security issues across 520 affected components. The primary vulnerability vector is debug logging, which accounts for approximately 73.5% of leaks, often exposing secrets directly into the LLM's context window. Most leaks occur during routine execution without requiring specific exploits, making them difficult for human oversight to detect. The study highlights that supply chain vulnerabilities persist, as secrets removed from upstream repositories often remain in forks. Experts recommend shifting from manual review to structural prevention, including the implementation of least-privilege access, credential rotation, log redaction, and robust sandboxing. The findings emphasize that treating AI agent extensions as untrusted code is essential for maintaining security in modern AI-driven development environments.