The Agent Skills Directory

[REMOTE_CODE_EXECUTION]: The skill documentation introduces the dspy.ProgramOfThought module, which is designed to generate and execute Python code at runtime to solve logic and math problems. This creates a primary attack surface where malicious input could influence the model to generate and execute unauthorized code.
[COMMAND_EXECUTION]: The agent implementation examples in references/examples.md include a calculate tool that utilizes the Python eval() function to process mathematical expressions. Although it attempts to restrict the environment by removing builtins, eval() is a high-risk primitive that can be exploited for arbitrary code execution.
[EXTERNAL_DOWNLOADS]: The installation instructions in SKILL.md suggest fetching the framework directly from a remote GitHub repository (github.com/stanfordnlp/dspy.git), which involves executing external code during the installation phase.
[PROMPT_INJECTION]: The skill's architecture exhibits a surface for indirect prompt injection due to its integration with external data sources and powerful execution capabilities. 1. Ingestion points: Data is ingested from Wikipedia summaries, web search results, and local document databases. 2. Boundary markers: The framework uses Signatures to structure inputs, but lacks explicit delimiters or instructions to ignore embedded commands in the retrieved data. 3. Capability inventory: The skill possesses the ability to execute generated Python scripts and use tools with expression evaluation access. 4. Sanitization: The provided examples do not demonstrate sanitization or filtering of external content before it is processed by the language model or execution modules.

dspy