The Agent Skills Directory

[DYNAMIC_EXECUTION]: The skill's documentation in SKILL.md and references/examples.md provides ReAct agent implementation examples that use the eval() function within a 'calculator' tool. This tool evaluates strings directly from the model's action_input, which allows for arbitrary code execution if the model is subverted via prompt injection.
[INDIRECT_PROMPT_INJECTION]: The skill demonstrates patterns for ingesting and processing untrusted external data (e.g., in extract_entities and classify_article) without demonstrating best practices for safety.
Ingestion points: Untrusted external strings are ingested through the text parameter in several extraction and classification functions.
Boundary markers: The examples lack explicit delimiters or specific instructions to the model to ignore potential commands embedded within the input text.
Capability inventory: The skill demonstrates the ability to execute code via eval() and perform complex multi-step reasoning, which increases the potential impact of an injection.
Sanitization: No input validation, filtering, or escaping techniques are shown for external data before it is interpolated into the model context.

guidance