The Agent Skills Directory

[DYNAMIC_EXECUTION]: The skill's primary purpose is the implementation of dspy.CodeAct, which allows an AI agent to dynamically generate and execute Python code at runtime. While this glue code is isolated in a Deno-based sandbox, the examples.md file provides an implementation of a mathematical compute tool that uses the eval() function, which is a common dynamic execution risk, even with a restricted global environment.
[INDIRECT_PROMPT_INJECTION]: The skill documents patterns for agents that ingest and process external data, such as CSV parsing and sales data analysis in examples.md. This establishes a vulnerability surface where malicious instructions embedded in the processed data could influence the agent's code generation process. The analysis identifies ingestion points in examples.md (FILE_STORE and SALES_DATA), notes the absence of explicit boundary markers in the code snippets, and identifies capabilities like state modification through tools such as write_result.
[COMMAND_EXECUTION]: The documentation explicitly states that tool functions execute in the normal Python process with full privileges. Examples provided in examples.md, such as the write_result function, demonstrate the ability to modify storage state based on agent-determined content, highlighting the importance of validating inputs at the trust boundary.

dspy-codeact