mcp-builder

CRITICAL E006: Malicious code pattern detected in skill scripts.

• Malicious code pattern detected (high risk: 0.90). The evaluation harness intentionally sends tool inputs and tool outputs to an external LLM (Anthropic) and explicitly instructs that the model must include the inputs/outputs in its , creating a high-risk data-exfiltration vector (sensitive tool results or environment-derived secrets could be sent to and recorded by the remote model service).

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

• Third-party content exposure detected (medium risk: 0.30). In scripts/evaluation.py, the evaluation harness ingests qa_pair text from an outsider-provided evaluation XML file at runtime and passes it directly as the LLM user message, enabling indirect prompt-injection via the attacker-authored question content.