The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill ingests and processes data from multiple external sources not under the direct control of the author, creating a potential surface for indirect prompt injection.
Ingestion points: Tool outputs from PubMed, CIViC, ClinVar, and ClinicalTrials.gov are integrated into the agent's context during the interpretation and recommendation phases.
Boundary markers: The skill does not employ explicit delimiters or system instructions for the agent to ignore or isolate potentially malicious instructions embedded in external content.
Capability inventory: The agent possesses capabilities for network operations across numerous APIs and the execution of a local Python script via subprocess.run.
Sanitization: There is no evidence of specific sanitization or validation of external tool outputs before they are processed into final narrative reports.
[EXTERNAL_DOWNLOADS]: The skill performs automated downloads of research data from a well-known scientific institution.
Evidence: The included script scripts/gdsc_drug_response.py utilizes urllib.request.urlretrieve to fetch Excel-formatted drug sensitivity datasets from the Wellcome Sanger Institute's domain (cog.sanger.ac.uk).
[COMMAND_EXECUTION]: The skill facilitates the invocation of a local helper script to process scientific data.
Evidence: The API_USAGE_PATTERNS.md file defines a pattern for executing scripts/gdsc_drug_response.py using subprocess.run with a list-based argument format, which mitigates common shell injection risks.

tooluniverse-precision-oncology