ai-sdk-core

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

• Third-party content exposure detected (high risk: 0.90). The skill explicitly instructs the agent to load and ingest external, potentially-untrusted content — e.g., createMCPClient() experimental_getPrompt and resources in references/mcp-integration.md, built-in web_fetch/web_search/urlContext tools and the researchTool/exaCrawl pattern in references/dynamic-tools.md — and shows injecting that fetched content into generateText prompts and middleware (ragMiddleware), meaning third-party pages/resources can directly influence tool use and agent decisions.

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

• Potentially malicious external URL detected (high risk: 0.90). The docs show runtime MCP client usage that fetches tools/prompts from an external HTTP MCP endpoint (e.g., https://your-server.com/mcp), and those fetched prompts/tools are then injected into generateText (experimental_getPrompt / client.tools()) so remote content can directly control agent prompts and tool execution.