meta-feasibility-analyzer
Pass
Audited by Gen Agent Trust Hub on Sep 28, 2026
Risk Level: SAFEPROMPT_INJECTIONINDIRECT_PROMPT_INJECTIONEXTERNAL_DOWNLOADS
Full Analysis
- [PROMPT_INJECTION]: The SKILL.md file defines workflows that interpolate user-provided titles ({{input_the_title}}) directly into high-level instructions for the LLM in Steps 1 and 8. This creates a surface where a malicious title could attempt to override the agent's behavior.
- [INDIRECT_PROMPT_INJECTION]: The skill processes data from the PubMed API which is then used as context for analysis by the LLM.
- Ingestion points: PubMed article metadata is retrieved in scripts/feasibility_ops.py using the Bio.Entrez module.
- Boundary markers: The LLM prompt in SKILL.md Step 8 does not use clear delimiters or negative constraints to prevent the agent from obeying instructions that might be embedded in retrieved article titles.
- Capability inventory: The skill allows script execution (feasibility_ops.py) and network requests to the NCBI PubMed service.
- Sanitization: The format_clinical and format_meta functions in scripts/feasibility_ops.py assemble summary strings from article titles without escaping or sanitizing the content before it is passed to the LLM.
- [EXTERNAL_DOWNLOADS]: The skill depends on the third-party biopython library. The SKILL.md file notes that dependencies are not version-pinned, which is a potential risk factor for supply chain integrity if the execution environment is not strictly controlled.
Audit Metadata