claude-code-agent-sdk

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

• Third-party content exposure detected (medium risk: 0.30). Outsider-authored free text is ingested as the runtime prompt/message stream provided to query() (and optionally via streaming input), before any agent skill selection can occur.

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

• Potentially malicious external URL detected (high risk: 0.90). The SDK example configures an MCP server with the runtime URL "https://code.claude.com/docs/mcp", which the agent will connect to at runtime to load tool definitions that can directly influence agent prompts/behavior.