The Agent Skills Directory

[DYNAMIC_EXECUTION]: The backend configuration guide in references/backends.md includes the trust_remote_code=True parameter when initializing models with vLLM. This is a standard configuration in the Hugging Face ecosystem that allows for the execution of custom architecture code defined within a model's repository. While this facilitates the use of community-contributed models, it represents a dynamic execution pathway for remote code provided by model authors.
[INDIRECT_PROMPT_INJECTION]: The skill is primarily designed to ingest and process external, untrusted text data to extract structured information into Pydantic models. Examples in SKILL.md and references/examples.md demonstrate the agent processing user-provided strings. This functionality introduces a surface for indirect prompt injection, where malicious instructions could be embedded in the input data. However, the skill's use of grammar-based generation (via Outlines) inherently limits the impact by forcing the LLM's output to conform strictly to a predefined JSON schema or Pydantic model.

outlines