blackbox-desktop

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 includes connecting to external MCP servers (e.g., via blackbox mcp add ... https://.../api/mcp), where the agent can ingest tool/resource outputs and remote content into its LLM context—an outsider-controlled source (public/third-party MCP server) that the user did not author.

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

• Potentially malicious external URL detected (high risk: 1.00). The skill explicitly instructs adding an MCP server at runtime using the URL https://cloud.blackbox.ai/api/mcp (via the blackbox mcp add ... command), and the docs state that connecting to that MCP exposes remote tools including "Remote code execution" and cloud-hosted agent tasks, which can directly control the agent's actions and execute code.