citation-integrity-reviewer

Warn

Audited by Gen Agent Trust Hub on Jul 26, 2026

Risk Level: MEDIUMCOMMAND_EXECUTIONPROMPT_INJECTION
Full Analysis
  • [COMMAND_EXECUTION]: The agent is directed to execute a local Python script using a JSON object derived from user-provided content. Evidence: 'Programmatically validate all DOIs... by executing python .codex/skills/scripts/validate_references.py "<doi_list_json>"' in SKILL.md. This pattern of passing AI-extracted strings directly into shell commands is vulnerable to injection attacks if the agent extracts malicious payload characters from the user's draft into the JSON argument.
  • [INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted drafts and bibliographies while maintaining script execution capabilities. Ingestion points: 'draft text' and 'source list' in SKILL.md. Boundary markers: Absent. Capability inventory: Command execution via Python. Sanitization: Absent. The skill lacks instructions to sanitize draft content or protect the execution environment from instructions embedded in the processed text.
Audit Metadata
Risk Level
MEDIUM
Analyzed
Jul 26, 2026, 09:51 AM
Security Audit — agent-trust-hub — citation-integrity-reviewer