The Agent Skills Directory

[DYNAMIC_EXECUTION]: The skill utilizes torch.compile (PyTorch 2.0+) which generates and compiles GPU kernels at runtime to optimize model performance. It also covers model deserialization using torch.load and torch.jit, identifying the risk of arbitrary code execution and recommending the use of weights_only=True.
[INDIRECT_PROMPT_INJECTION]: The skill defines a surface for indirect prompt injection. 1. Ingestion points: CustomDataset in SKILL.md allows for arbitrary data processing. 2. Boundary markers: No explicit markers are provided to delimit external data from agent instructions. 3. Capability inventory: Includes dynamic code generation via torch.compile and file system writes via torch.save and torch.cuda.memory._dump_snapshot. 4. Sanitization: No input validation or sanitization is implemented for the data loaders.

pytorch