youtube-thumbnail

Fail

Audited by Gen Agent Trust Hub on Feb 15, 2026

Risk Level: HIGHEXTERNAL_DOWNLOADSREMOTE_CODE_EXECUTIONPROMPT_INJECTIONCOMMAND_EXECUTION
Full Analysis
  • EXTERNAL_DOWNLOADS (HIGH): The skill requires the installation of the thumbkit tool from a non-trusted GitHub repository (https://github.com/kenneth-liao/thumbkit.git).\n- REMOTE_CODE_EXECUTION (HIGH): Executing uv tool install on an untrusted repository allows for the execution of arbitrary third-party code during installation or runtime. Evidence: uv tool install https://github.com/kenneth-liao/thumbkit.git in SKILL.md.\n- PROMPT_INJECTION (HIGH): The skill is vulnerable to indirect prompt injection through the ingestion of external images. Attackers can embed malicious instructions in YouTube thumbnails or logos which the agent is then instructed to download and 'read' to understand. \n
  • Ingestion points: External image URLs from web searches and YouTube thumbnail URLs (SKILL.md).\n
  • Boundary markers: Absent. There are no instructions to delimit or ignore content within the downloaded files.\n
  • Capability inventory: Shell command execution via uv, curl, and the thumbkit CLI tool.\n
  • Sanitization: Absent. No validation or sanitization is performed on the downloaded content before processing.\n- COMMAND_EXECUTION (MEDIUM): The skill uses curl to download assets from arbitrary, potentially attacker-controlled URLs found during search operations.
Recommendations
  • AI detected serious security threats
Audit Metadata
Risk Level
HIGH
Analyzed
Feb 15, 2026, 10:36 PM