The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and analyze untrusted AI model files which may contain embedded instructions or malicious payloads. Ingestion points include model loading functions described in SKILL.md such as onnx.load and torch.load. Boundary markers are present in the form of strong, explicit warnings against loading unverified Pickle-based files. The capability inventory includes local file analysis, weight extraction to .npy files, and hash calculation. Sanitization is addressed by recommending the use of weights_only=True for PyTorch loading and the transition to safer formats like Safetensors.
[DYNAMIC_EXECUTION]: The skill documents the specific risks of arbitrary code execution inherent in the PyTorch Pickle format. It provides technical evidence of how torch.load can be exploited via Python's Pickle protocol and instructs the user on defensive programming techniques to mitigate these risks, such as using isolated sandbox environments for unknown models.
[EXTERNAL_DOWNLOADS]: The skill provides setup instructions for industry-standard AI analysis tools and libraries. It includes commands for pip install targeting packages like torch and onnx from official registries, and installation commands for system packages via trusted managers like apt and brew. These references target well-known services and official project domains.
[COMMAND_EXECUTION]: The skill utilizes local system utilities to facilitate the forensic analysis of binary files. This includes the use of sha256sum for integrity verification, xxd for hex inspection, and netron for graph visualization. These operations are performed locally on the analysis target and are standard for reverse engineering workflows.

re-ai-model