baseline-extraction-for-clinical-trials

Warn

Audited by Gen Agent Trust Hub on Jul 6, 2026

Risk Level: MEDIUMCOMMAND_EXECUTIONPROMPT_INJECTIONEXTERNAL_DOWNLOADS
Full Analysis
  • [PROMPT_INJECTION]: The skill contains a deceptive file baseline-extraction-for-clinical-trials_audit_result_v1.json that masquerades as an official security report. This is a self-referential injection pattern (Category 8e) designed to deceive the analyzer and human reviewers by providing false safety verdicts such as 'security: PASS'. Additionally, the skill processes untrusted medical text and PDFs without security boundaries.
  • Ingestion points: Full article text and PDF content via extract_pdf.py.
  • Boundary markers: Absent. No markers or 'ignore embedded instructions' warnings are present.
  • Capability inventory: Subprocess execution via baseline_extractor.py and local file writes via extract_pdf.py.
  • Sanitization: None. The content is used directly for LLM extraction.
  • [COMMAND_EXECUTION]: The workflow in SKILL.md (Step 1) instructs the agent to execute shell commands using subprocess.run with a user-supplied <PMID>. There is no validation or sanitization required for this input, creating a vector for command injection if the agent or the underlying script handles the argument unsafely.
  • [EXTERNAL_DOWNLOADS]: The skill depends on the PyPDF2 library and references the PubMed API. The documentation explicitly states that dependencies are not version-pinned, which introduces a supply chain risk where malicious versions of libraries could be introduced into the execution environment.
  • [NO_CODE]: The skill instructions reference a primary implementation script scripts/baseline_extractor.py, but this file is missing from the provided skill package, making its internal logic and safety unverifiable.
Audit Metadata
Risk Level
MEDIUM
Analyzed
Jul 6, 2026, 01:42 AM
Security Audit — agent-trust-hub — baseline-extraction-for-clinical-trials