The Agent Skills Directory

[DYNAMIC_EXECUTION]: The documentation for DSPy promotes the use of dspy.ProgramOfThought, which generates and executes Python code at runtime to perform reasoning and calculations. This behavior can lead to arbitrary code execution if the module processes untrusted input strings.\n- [COMMAND_EXECUTION]: In references/examples.md, the skill demonstrates a calculation tool implementation using the Python eval() function: result = eval(expression, {"__builtins__": {}}, {}). Using eval on string input is a high-risk pattern that can often be bypassed to execute unauthorized commands, even when built-ins are restricted.\n- [INDIRECT_PROMPT_INJECTION]: The skill represents a framework for ingesting external data through retrieval and user inputs, creating a surface for indirect injection attacks.\n
Ingestion points: Data enters the system via dspy.InputField() definitions in SKILL.md and references/examples.md.\n
Boundary markers: The examples lack explicit delimiters or instructions to ignore malicious commands embedded in processed data.\n
Capability inventory: The skill provides access to code execution (ProgramOfThought) and tool-using agents (ReAct) that can perform arithmetic or search operations.\n
Sanitization: There is no evidence of sanitization or strict schema validation for input fields before they are used in prompt generation.\n- [METADATA_POISONING]: The SKILL.md metadata lists the author as 'Orchestra Research', which is inconsistent with the expected author 'firecrawl', potentially misleading users about the skill's origin and maintenance.

dspy