dspy
Pass
Audited by Gen Agent Trust Hub on Sep 15, 2026
Risk Level: SAFEEXTERNAL_DOWNLOADSCOMMAND_EXECUTIONDYNAMIC_EXECUTIONINDIRECT_PROMPT_INJECTIONMETADATA_POISONING
Full Analysis
- [EXTERNAL_DOWNLOADS]: The skill references the official GitHub repository for the DSPy library (github.com/stanfordnlp/dspy.git), which is a well-known project from the Stanford NLP Group.
- [COMMAND_EXECUTION]: Provides instructions for installing required Python dependencies using pip, including dspy, openai, and anthropic.
- [DYNAMIC_EXECUTION]: Documents the dspy.ProgramOfThought module, which generates and executes Python code to solve math and logic problems. It also includes a calculate tool example in references/examples.md that utilizes a sandboxed eval() call.
- [INDIRECT_PROMPT_INJECTION]: The skill demonstrates building AI systems that process external user inputs and interpolate them into prompts. 1. Ingestion points: User-provided inputs in Predict, ChainOfThought, and ReAct modules (e.g., question field in SKILL.md). 2. Boundary markers: No specific delimiters or warnings are included in the provided prompt templates. 3. Capability inventory: The systems have access to tool execution via ReAct and dynamic code execution via ProgramOfThought. 4. Sanitization: No input sanitization or validation logic is shown in the examples.
- [METADATA_POISONING]: The author field in SKILL.md ('Orchestra Research') differs from the system-reported author ('davila7'), although this appears to be an organizational identifier rather than a malicious impersonation.
Audit Metadata