using-tensor-compiler-engineering

Pass

Audited by Gen Agent Trust Hub on Aug 10, 2026

Risk Level: SAFEPROMPT_INJECTION
Full Analysis
  • [PROMPT_INJECTION]: The skill is designed to ingest and process Intermediate Representations (IR) of tensor programs from various sources (DSLs, ONNX, etc.), as described in ir-contracts-and-semantic-identity.md. This represents a surface for Indirect Prompt Injection where malicious logic could be embedded in the graph structure to influence the compiler's output or the agent's behavior.
  • Ingestion points: The 'Ingest' stage in compiler-architecture-for-tensor-programs.md processes external IR graphs.
  • Boundary markers: The skill emphasizes 'IR Contracts' and semantic hashes for identity preservation, but lacks explicit security boundaries to isolate or sanitize adversarial graph inputs.
  • Capability inventory: The skill utilizes torch.compile, torch.fx, and AOTAutograd to perform runtime code generation and execution of artifacts.
  • Sanitization: No security-focused sanitization of the graph structure is implemented beyond semantic validation.
  • [COMMAND_EXECUTION]: The skill leverages the PyTorch compilation stack (Dynamo, AOTAutograd, Inductor) to perform dynamic code generation and execution, as detailed in torch-compile-and-aotautograd.md. While these are standard tools for tensor compilation, they represent a significant capability tier that requires carefully controlled inputs.
Audit Metadata
Risk Level
SAFE
Analyzed
Aug 10, 2026, 08:34 PM
Security Audit — agent-trust-hub — using-tensor-compiler-engineering