The Agent Skills Directory

[DYNAMIC_EXECUTION]: The skill documentation includes code examples that utilize the eval() function to process mathematical expressions. Specifically, references/modules.md contains a tool definition that performs an unrestricted eval(expression), which could lead to arbitrary code execution if the input is manipulated by an LLM influenced by external prompts. Another example in references/examples.md uses a restricted environment for eval(), but the practice is consistently promoted without sufficient warnings.
[DYNAMIC_EXECUTION]: The skill prominently features the dspy.ProgramOfThought module, which is designed to generate and execute Python code at runtime based on natural language inputs, inherently incorporating model-generated content into the execution flow.
[INDIRECT_PROMPT_INJECTION]: The skill provides templates for building RAG (Retrieval-Augmented Generation) and agent systems that ingest untrusted data from external sources such as search results or user-provided documents. These templates do not include explicit sanitization or validation logic, creating a surface for indirect prompt injection attacks. 1. Ingestion points: Untrusted data enters the context via the question, text, and document parameters in various module examples in SKILL.md and references/examples.md. 2. Boundary markers: The examples lack boundary markers or instructions to the model to ignore embedded commands within the processed data. 3. Capability inventory: The modules are equipped with code execution capabilities via ProgramOfThought and the ability to use external tools through ReAct agents. 4. Sanitization: No sanitization, escaping, or filtering of retrieved content is shown in the pipeline implementations.
[EXTERNAL_DOWNLOADS]: The installation instructions suggest installing the framework directly from a remote GitHub repository (pip install git+https://github.com/stanfordnlp/dspy.git). This is documented as the official source for the library.

dspy