clinical-decision-support
Warn
Audited by Gen Agent Trust Hub on Sep 19, 2026
Risk Level: MEDIUMCOMMAND_EXECUTIONINDIRECT_PROMPT_INJECTIONDYNAMIC_EXECUTION
Full Analysis
- [COMMAND_EXECUTION]: The skill provides a template for schematic generation:
python scripts/generate_schematic.py "your diagram description" -o figures/output.png. This pattern instructs the agent to interpolate a natural language description directly into a shell command, which is susceptible to command injection if the user-provided description contains shell metacharacters such as backticks or subshells. - [INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and process untrusted clinical data, trial results, and patient cohort descriptions. \n 1. Ingestion points: User-provided clinical descriptions and cohort data (e.g., NSCLC patient cohorts, GBM molecular subtypes). \n 2. Boundary markers: None are defined in the instructions to prevent the agent from obeying instructions embedded in the processed data. \n 3. Capability inventory: The skill has
Bash(script execution),Write(file creation), andEditpermissions. \n 4. Sanitization: There are no instructions to sanitize or escape data before it is inserted into LaTeX templates, which could lead to LaTeX injection or adversarial influence over the agent's output. - [DYNAMIC_EXECUTION]: The skill utilizes a suite of local scripts (
scripts/*.py) to perform complex tasks like survival analysis (generate_survival_analysis.py) and biomarker classification at runtime. While these scripts are local, their execution with user-controlled input data introduces a surface for runtime execution risks.
Audit Metadata