The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill instructs the user to clone executable code from an external GitHub repository (github.com/tonbistudio/turboquant-pytorch).\n- [REMOTE_CODE_EXECUTION]: The instructions direct the user to execute the downloaded code via shell commands such as python -m turboquant.test_turboquant and python -m turboquant.validate.\n- [DYNAMIC_EXECUTION]: The troubleshooting section recommends using the pickle module to save and load codebooks. The pickle module is insecure and can be exploited to execute arbitrary code if it loads a maliciously crafted file.\n- [INDIRECT_PROMPT_INJECTION]: The skill fetches pre-trained models (e.g., Qwen2.5-3B-Instruct) from the Hugging Face hub during validation.\n
Ingestion points: External model files are downloaded and processed by validation scripts.\n
Boundary markers: No explicit delimiters or instructions to ignore embedded content are present in the model loading process.\n
Capability inventory: The skill environment allows command execution (git, pip, python), file system writes for caching, and unsafe file loading via pickle.\n
Sanitization: There are no integrity checks or sanitization routines mentioned for the downloaded model data or the serialized codebook files.

turboquant-pytorch