karpathy-wiki-capture

Warn

Audited by Gen Agent Trust Hub on May 11, 2026

Risk Level: MEDIUMCOMMAND_EXECUTIONPROMPT_INJECTION
Full Analysis
  • [COMMAND_EXECUTION]: The skill relies on executing local shell commands (bin/wiki, mv, ls, grep) to perform its core functions. It instructs the agent to assemble command strings using variables derived from conversation or user input.
  • Evidence: echo "$BODY" | bin/wiki capture --title "<one-line title>" in SKILL.md.
  • Evidence: mv <subagent-report-path> <wiki>/inbox/<basename> in SKILL.md.
  • [PROMPT_INJECTION]: The skill is vulnerable to indirect prompt injection as it ingests and processes untrusted data from both user conversations and subagent report files.
  • Ingestion points: The $BODY variable (derived from chat) and external report files specified by <subagent-report-path> (found in SKILL.md).
  • Boundary markers: No specific delimiters or instructions to ignore embedded instructions are provided in the command templates.
  • Capability inventory: The skill has the ability to execute shell commands (bin/wiki), move files (mv), and list directory contents (ls).
  • Sanitization: There are no instructions or automated steps to sanitize the input data before it is passed to the shell or written to the wiki infrastructure.
Audit Metadata
Risk Level
MEDIUM
Analyzed
May 11, 2026, 11:31 AM
Security Audit — agent-trust-hub — karpathy-wiki-capture