skills/aipoch/medical-research-skills/quadas-c-assessment-for-diagnostic-accuracy-studies/Gen Agent Trust Hub
quadas-c-assessment-for-diagnostic-accuracy-studies
Warn
Audited by Gen Agent Trust Hub on Sep 28, 2026
Risk Level: MEDIUMMETADATA_POISONINGINDIRECT_PROMPT_INJECTION
Full Analysis
- [METADATA_POISONING]: Deceptive Audit Metadata. The skill includes a file named 'quadas-c-assessment-for-diagnostic-accuracy-studies_audit_result_v1.json' which is a fabricated security audit record. This file contains fields like 'security: PASS' and 'deployable: true' intended to impersonate a formal safety review and mislead security analyzers or users about the skill's risk profile.
- [INDIRECT_PROMPT_INJECTION]: Vulnerable Data Ingestion surface. The skill processes untrusted external content from research papers and incorporates it into LLM prompts without adequate protection.
- Ingestion points: 'scripts/extract_pdf.py' (extracts text from user-provided PDFs) and 'scripts/quadas_c.py' (reads text from files or command-line arguments).
- Boundary markers: Absent. The scripts insert extracted text directly into prompt templates without using delimiters, XML-style tags, or instructions to ignore embedded commands.
- Capability inventory: The skill uses an LLM to identify diagnostic methods and perform domain-specific bias assessments, outputting the results as a JSON structure.
- Sanitization: Absent. The input text is not sanitized, escaped, or filtered before being processed by the LLM.
Audit Metadata