The Agent Skills Directory

[DYNAMIC_EXECUTION]: The file 'references/checkpoint-spec-pairing.md' recommends a Python snippet using torch.load(ckpt, map_location="cpu", weights_only=False) to derive architecture parameters from a model checkpoint. Using weights_only=False enables unsafe deserialization via the pickle module, which can execute arbitrary code embedded in a malicious checkpoint file.- [EXTERNAL_DOWNLOADS]: The skill instructs the user to download a model checkpoint from an external HuggingFace repository (zongzhuofan/co-detr-vit-large-coco) using the huggingface-cli. While HuggingFace is a well-known service, the repository is not owned by the primary vendor.- [COMMAND_EXECUTION]: The skill utilizes docker run to execute the codetr CLI for various tasks. It also employs shell logic to probe the container environment for the existence of specific scripts or Python modules at runtime.- [INDIRECT_PROMPT_INJECTION]: The skill is vulnerable to indirect prompt injection due to its processing of external data sources.
Ingestion points: The skill reads external data from COCO JSON files, image directories, and classmap text files.
Boundary markers: No explicit delimiters or instructions to ignore embedded commands are present in the processing workflow.
Capability inventory: The skill possesses the ability to execute shell commands within Docker containers and write files to the results directory.
Sanitization: There is no evidence of sanitization or strict schema validation for the input data before it is processed by the TAO entrypoints.

tao-train-codetr