dspy-evaluate

Pass

Audited by Gen Agent Trust Hub on Mar 22, 2026

Risk Level: SAFEPROMPT_INJECTIONCOMMAND_EXECUTION
Full Analysis
  • [PROMPT_INJECTION]: The skill demonstrates LM-as-judge patterns in examples.md (e.g., lines 88, 140, 149) that interpolate potentially untrusted data into evaluation prompts, creating a surface for indirect prompt injection attacks. \n
  • Ingestion points: Fields such as predicted_explanation, predicted_answer, and answer (from prediction) are fed directly into dspy.Signature classes. \n
  • Boundary markers: Absent. The signatures do not use delimiters (like triple quotes) or specific isolation instructions to distinguish untrusted content from the judge's instructions. \n
  • Capability inventory: The skill performs automated scoring and includes a CLI script for execution; scripts/run_eval.py can load program states from local files. \n
  • Sanitization: Absent. No evidence of input filtering or escaping is demonstrated in the provided examples. \n- [COMMAND_EXECUTION]: The scripts/run_eval.py utility restores program state from a file path provided via command-line arguments using dspy.Module().load(). While standard for the framework, loading serialized model states carries a risk of arbitrary code execution if the source file is malicious or if the underlying framework uses unsafe deserialization methods.
Audit Metadata
Risk Level
SAFE
Analyzed
Mar 22, 2026, 01:47 AM
Security Audit — agent-trust-hub — dspy-evaluate