ai-agent-observability-evaluation

MEDIUM W012: Unverifiable external dependency detected (runtime URL that controls agent).

• Potentially malicious external URL detected (high risk: 0.80). The skill loads judge prompts at runtime from the prompt registry (e.g., prompt://judges/support-resolution.v3 and prompt://judges/log-investigator-resolution.v2 via prompt_registry.get(contract.judge_prompt)), which directly supplies the LLM's instructions and is relied on by the judge cascade, so this is a runtime external dependency that controls prompts.