The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill is designed to process external data (text, images, and audio) using pre-trained models. Malicious content within these inputs could potentially influence the model's output to include instructions that an agent might follow, leading to unintended actions.
Ingestion points: Several examples in SKILL.md (Quick Start), references/pipelines.md, and references/training.md show the ingestion of external data for inference and training.
Boundary markers: No explicit delimiters or instructions to ignore embedded commands are present in the provided templates.
Capability inventory: The skill is granted Bash, Write, and Edit permissions in SKILL.md, which increases the potential impact if the agent executes instructions found in processed data.
Sanitization: There is no documentation or evidence of input/output sanitization or validation logic.
[DYNAMIC_EXECUTION]: The documentation references the trust_remote_code=True parameter, which allows the execution of arbitrary Python code from a model's repository on the Hugging Face Hub.
Evidence: SKILL.md and references/models.md include instructions on when to use this parameter.
Mitigation: The skill includes a specific security warning in SKILL.md, advising users to only use this feature after reviewing the custom code, which demonstrates proactive safety guidance.

transformers