dspy-gepa-optimizer
Pass
Audited by Gen Agent Trust Hub on Sep 14, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTION
Full Analysis
- [SAFE]: The skill provides documentation and boilerplate code for prompt optimization using the DSPy framework. Analysis of the Python examples and Markdown documentation shows adherence to standard library usage with no evidence of malicious intent. No suspicious network activities, unauthorized file access, or obfuscation techniques were found.
- [INDIRECT_PROMPT_INJECTION]: The skill processes user-provided datasets (trainset, valset) through an LLM-based optimizer (dspy.GEPA). This represents an indirect prompt injection surface as untrusted data is used to generate instructions. 1. Ingestion points: Training and validation datasets passed to optimizer.compile() in example_gepa.py and example_bettertogether.py. 2. Boundary markers: None explicitly defined in the provided code; relies on default DSPy prompt templates. 3. Capability inventory: File system writes (optimized_program.json via optimized.save()) and network API calls to LLM providers (via dspy.LM). 4. Sanitization: None; raw examples are used to generate natural language feedback. The severity is assessed as safe as this behavior is inherent to the primary intended purpose of the prompt optimizer skill.
Audit Metadata