ilya-sutskever-perspective

Warn

Audited by Gen Agent Trust Hub on Oct 13, 2026

Risk Level: MEDIUMCOMMAND_EXECUTIONINDIRECT_PROMPT_INJECTIONEXTERNAL_DOWNLOADS
Full Analysis
  • [COMMAND_EXECUTION]: The skill instructs the agent to perform automated tasks involving shell command execution and file operations.
  • Evidence in SKILL.md (版本自检 section): The agent is directed to read and write a local file .last-update-check and execute git commands (e.g., git rev-parse HEAD, git ls-remote origin HEAD) to check for updates periodically.
  • This mechanism enables the triggering of network and filesystem operations through the agent based on skill logic without per-invocation user confirmation.
  • [INDIRECT_PROMPT_INJECTION]: The skill implements a workflow that mandates the ingestion of untrusted external data.
  • Evidence in SKILL.md (Step 2: Ilya式研究): The agent is instructed to use WebSearch tools to gather technical information, papers, and news from the public internet.
  • The ingestion of this untrusted data, combined with the agent's capability to execute commands and perform further searches, creates a surface for indirect prompt injection as there are no specified boundary markers or sanitization instructions for the fetched content.
  • [EXTERNAL_DOWNLOADS]: The skill references and promotes the installation of external code and tools from remote sources.
  • Evidence in README.md: Installation instructions using npx skills add for author-specific packages and community tools (e.g., alchaincyf/ilya-sutskever-skill, vercel-labs/skills).
  • Evidence in README.md and SKILL.md: References to multiple repositories on GitHub for cloning and automated version tracking.
Audit Metadata
Risk Level
MEDIUM
Analyzed
Oct 13, 2026, 10:04 PM