The Agent Skills Directory

[DYNAMIC_EXECUTION]: The skill provides multiple implementation examples (found in SKILL.md and references/examples.md) that utilize the Python eval() function to process strings generated by a Large Language Model. Specifically, the 'ReAct Agent' examples define a calculator tool as lambda expr: eval(expr), where expr is directly sourced from model generation via gen('action_input'). This pattern is high-risk as it allows for arbitrary code execution on the host system if the model is influenced by malicious input.\n- [INDIRECT_PROMPT_INJECTION]: The skill documents architectural patterns for building agents that ingest untrusted external data and pass it into templates with high-privilege execution capabilities like eval().\n
Ingestion points: Agent templates in SKILL.md and references/examples.md ingest raw text through parameters like question and text.\n
Boundary markers: The provided examples do not demonstrate the use of delimiters or specific safety instructions to prevent the agent from executing commands embedded within the processed external data.\n
Capability inventory: The documented templates include an eval()-based calculator tool, providing an exploitable capability surface for any instructions extracted from the input.\n
Sanitization: The documentation does not demonstrate sanitization or strict validation of the input data or the resulting model-generated tool arguments before they reach the eval() function.\n- [METADATA_POISONING]: The skill exhibits inconsistent authorship metadata, declaring 'Orchestra Research' in the YAML frontmatter of SKILL.md while being associated with the author 'qcmuu' in the skill's deployment context.

guidance