tooluniverse-drug-target-validation
Pass
Audited by Gen Agent Trust Hub on Oct 7, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTIONDYNAMIC_EXECUTIONCOMMAND_EXECUTION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill processes information retrieved from external databases, creating a vector for potential adversarial instructions hidden in biological data or literature.\n
- Ingestion points: Data from tools like PubMed, OpenTargets, ChEMBL, and user-supplied target names are used throughout the 10-phase pipeline.\n
- Boundary markers: While the skill emphasizes source citation, it does not implement explicit delimiters or safety warnings for the processing of external content.\n
- Capability inventory: The agent has access to a Bash environment and the ability to execute Python scripts.\n
- Sanitization: The instructions do not define methods for escaping or validating the content retrieved from external biological tools.\n- [DYNAMIC_EXECUTION]: The skill relies on the runtime generation and execution of Python code to perform statistical calculations and scoring.\n
- Evidence: The 'COMPUTE, DON'T DESCRIBE' directive and the provided ADMET-AI fallback code block encourage the creation and execution of scripts at runtime.\n- [COMMAND_EXECUTION]: The pipeline utilizes the system shell to execute Python-based analysis and invoke various biological informatics tools.
Audit Metadata