torch-geometric
Pass
Audited by Gen Agent Trust Hub on Oct 4, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTION
Full Analysis
- [SAFE]: The skill maintains a clear 'SandBase boundary', ensuring that core PyTorch Geometric data processing remains local unless the user explicitly requests external datasets or model inference. It mandates user confirmation for paid calls, requires price transparency, and emphasizes minimal data transfer.
- [INDIRECT_PROMPT_INJECTION]: The skill possesses a surface for indirect prompt injection as it is designed to ingest data from external sources and use generative models through the SandBase API. This surface is mitigated by the mandatory implementation of validation workflows.
- Ingestion points: External datasets and model inference results obtained through the sandbase_run tool (SKILL.md, references/sandbase-api-map.md).
- Boundary markers: Instructions requiring the agent to restate requested outcomes, constraints, and assumptions before implementation (Workflow step 2, SKILL.md).
- Capability inventory: Local graph neural network implementation, data analysis, and artifact generation (Workflow step 3, SKILL.md).
- Sanitization: Mandatory quality gates for inspecting data shapes, types, units, and train/test boundaries, as well as focus checks for graph features and label leakage (Quality gates/Focus checks, SKILL.md).
Audit Metadata