The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill facilitates the processing of model artifacts like .pt and .pth files, which could potentially contain malicious content if sourced from untrusted environments.
Ingestion Points: The skill instructs the agent to load files using torch.load("model.pt", ...) and torch.load("weights.pth", ...).
Boundary Markers: The instructions explicitly warn to use weights_only=True to mitigate risks associated with untrusted artifacts.
Capability Inventory: The skill involves reading files and materializing tensors in memory.
Sanitization: The skill advocates for specific sanitization techniques including safe_globals and restricted unpickling to prevent arbitrary code execution during the loading process.
[DYNAMIC_EXECUTION]: The skill documents the use of torch.load, which utilizes the pickle serialization format.
Evidence: The skill notes that torch.load can allow arbitrary code execution if not configured correctly (using weights_only=False).
Mitigation: The skill provides extensive documentation and code examples on how to use weights_only=True and add_safe_globals to avoid the dangers of dynamic code execution when deserializing model data.

pytorch