openai-responses

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

• Third-party content exposure detected (high risk: 1.00). The skill explicitly uses a built-in "web_search" tool (see references/built-in-tools-guide.md and multiple templates like templates/cloudflare-worker.ts and background-mode.ts) and MCP integrations that fetch real-time content from public websites and external MCP servers, so the agent will ingest and interpret untrusted third-party web content as part of its workflow.

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

• Potentially malicious external URL detected (high risk: 0.90). The skill includes runtime MCP server URLs (e.g., https://mcp.stripe.com) passed to the Responses API as tools; the model will contact those servers to discover and invoke tools, allowing remote endpoints to directly control agent behavior and execute actions.

MEDIUM W009: Direct money access capability detected (payment gateways, crypto, banking).

• Direct money access detected (high risk: 1.00). The skill explicitly includes MCP (Model Context Protocol) integration for external tools and gives a concrete example using Stripe: it shows an MCP tool entry with server_label 'stripe', server_url 'https://mcp.stripe.com', and an authorization token, and an example input "Create a $20 payment link" with the comment "Model uses Stripe MCP server to create payment link." This is a specific payment-gateway integration (Stripe) and demonstrates creating a payment action, which qualifies as Direct Financial Execution.