research-question

Warn

Audited by Gen Agent Trust Hub on Jun 23, 2026

Risk Level: MEDIUMCOMMAND_EXECUTIONREMOTE_CODE_EXECUTIONPROMPT_INJECTION
Full Analysis
  • [COMMAND_EXECUTION]: The skill instructions define a command-line pattern <lit> search "<q>" where research questions (derived from user input or generated) are interpolated directly into shell commands. This creates a risk of command injection if the input contains shell metacharacters or quotes.\n- [REMOTE_CODE_EXECUTION]: The skill relies on an external Python script (lit_search.py) located in a configurable path (~/.claude/skills/literature-search/). The skill executes this script using python3, which represents the execution of code outside the skill's own package. While described as a 'sibling skill', this dependency is not verified and is controlled by a local configuration file (loop.run.yaml).\n- [PROMPT_INJECTION]: The skill possesses an indirect prompt injection surface by ingesting untrusted data from WebSearch, WebFetch, and the output of the literature search tool. This content is passed to a 'Grader' subagent to influence the scoring and refinement of questions.\n
  • Ingestion points: SKILL.md
  • 'Gather novelty evidence' step (WebSearch, WebFetch, LitSearch output).\n
  • Boundary markers: Absent. The instructions do not specify any delimiters or warnings to ignore embedded instructions in the search results.\n
  • Capability inventory: Subprocess execution of Python scripts, file writes to the sandbox and config files, and network operations via web search tools.\n
  • Sanitization: Absent. No explicit validation or escaping of external content before processing by the subagent is mentioned.
Audit Metadata
Risk Level
MEDIUM
Analyzed
Jun 23, 2026, 04:42 PM
Security Audit — agent-trust-hub — research-question