newsfilter
← Interviews
AI Engineer15 min

You Can't Tell What an Agent Skill Will Do by Reading It - Roey Ben Chaim, Zenity

Roey Ben Chaim

Conference Presentation, Keynote
  • Malicious actors may utilize scripts to download additional payloads or packages, potentially bypassing static scanners that rely solely on reviewing skills.md files.
  • LLM analyzers can be deceived by manipulation tactics such as falsified AppSec approval comments, and generic task prompts may fail to cover all necessary operational modes for diverse skills.
  • Risks include malicious routes within a skill trajectory, domain takeover, and the ability for threat actors to alter prompts over time, which could lead to major consequences if hooks like "user prompt submit" are triggered repeatedly.
  • Post-installation scripts (e.g., npm install) can invoke new processes to read/write files and connect to external endpoints, allowing attackers to mimic official tools and trick creators into installing malicious replacements.
  • Specific attack demonstrations involved fake organizations mimicking the Paperclip project, with Python and JavaScript variants executing credential harvesting, payload decoding, and self-destruction via control endpoints.
  • A coordinated incident involving shared command-and-control infrastructure resulted in 1.7 million downloads for affected skills before the threat actor's control and the official projects were taken down within the last 12 hours.
  • All data entering the context faces risk, including websites, database rows, packages, and comments, while safeguards may currently be inactive on scenarios that are definitively not malicious.
  • Testing capabilities for these threats are available via the detonation chamber tool on ARToto.io.