pytensor-workflows
Pass
Audited by Gen Agent Trust Hub on Sep 21, 2026
Risk Level: SAFEDYNAMIC_EXECUTIONINDIRECT_PROMPT_INJECTION
Full Analysis
- [SAFE]: The skill contains technical documentation and local Python scripts for PyTensor workflows. No malicious patterns such as prompt injection, data exfiltration, or unauthorized network operations were detected.\n- [DYNAMIC_EXECUTION]: The skill documents and uses PyTensor's compilation capabilities (to C, JAX, Numba, etc.). This is the primary intended function of the library. The documentation correctly identifies security risks associated with certain PyTensor features, such as the use of pickle in function_dump, and provides appropriate warnings to the user to avoid unpickling untrusted files.\n- [INDIRECT_PROMPT_INJECTION]: The skill instructs the agent on how to process symbolic mathematical expressions provided by users. \n
- Ingestion points: User-provided mathematical definitions and data arrays (e.g., in the minimal runnable graph example).\n
- Boundary markers: The instructions emphasize establishing explicit shape and type contracts (e.g., pt.vector("x", dtype="float64")) before execution to delimit valid input.\n
- Capability inventory: The skill uses pytensor.function to compile and execute symbolic graphs, and involves local script execution for custom Ops.\n
- Sanitization: PyTensor performs internal shape, rank, and dtype validation as described in references/shapes.md and the custom Op implementations.
Audit Metadata