azure-ai-fine-tuning

MEDIUM W011: Third-party content exposure detected (indirect prompt injection risk).

• Third-party content exposure detected (high risk: 0.75). The skill’s runtime workflow can ingest outsider-authored free text via the Foundry Data Generation API “traces” source: TracesDataGenerationJobSource reconstructs conversations from Application Insights (including user/customer messages and tool outputs) and the resulting generated JSONL is then downloaded and fed into the fine-tuning/eval pipeline as LLM-readable text.

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

• Potentially malicious external URL detected (high risk: 0.90). The skill includes agentic RFT/tool-calling and endpoint-grader examples that invoke external HTTP endpoints at runtime—e.g., tools with server_url "https://your-function-app.azurewebsites.net/api/tools" and an endpoint grader "https://your-grading-endpoint.com/grade"—which will be called during training/post-training to execute logic and return scores, so these are runtime external dependencies that execute remote code and control training behavior.