equation-discovery
Pass
Audited by Gen Agent Trust Hub on Nov 7, 2026
Risk Level: SAFEDYNAMIC_EXECUTIONCOMMAND_EXECUTIONEXTERNAL_DOWNLOADSINDIRECT_PROMPT_INJECTION
Full Analysis
- [DYNAMIC_EXECUTION]: The skill evaluates Python code skeletons generated by the agent during the discovery loop. To prevent the execution of malicious code, the
evaluatefunction inscripts/eqdisc.pyperforms static validation using theastmodule, which effectively blocks imports, file I/O operations, and dangerous built-in functions. The validated code is then executed within a restricted namespace containing only necessary libraries likenumpy.\n- [COMMAND_EXECUTION]: The evaluation of equation skeletons is performed by spawning a separate worker process usingsubprocess.runinscripts/eqdisc.py. This process is launched with the Python-I(isolated) flag to ensure it runs in a clean environment without access to user site-packages or environment-specific configurations.\n- [EXTERNAL_DOWNLOADS]: Thefetchcommand inscripts/eqdisc.pydownloads scientific benchmark datasets in CSV format from the official GitHub repository of the research group associated with the cited papers. This download targets a well-known research organization and is consistent with the skill's intended scientific purpose.\n- [INDIRECT_PROMPT_INJECTION]: The skill processes external CSV data, which represents an attack surface for indirect prompt injection. \n - Ingestion points: Data is loaded from CSV files using
pandas.read_csvinscripts/eqdisc.py.\n - Boundary markers: The skill employs AST-based code validation to ensure the processing logic remains restricted to mathematical operations.\n
- Capability inventory: The skill utilizes subprocess isolation and restricted execution namespaces for all data-driven calculations.\n
- Sanitization: Data is treated as numerical input for fitting algorithms, and the execution environment for skeletons is strictly sanitized through whitelist-based built-in filtering.
Audit Metadata