web-clone

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

• Third-party content exposure detected (high risk: 1.00). The skill's Pass 1 "Extract" workflow in SKILL.md explicitly instructs the agent to browser_navigate to a user-provided target_url and ingest the live page via browser_snapshot, browser_take_screenshot, and browser_evaluate (and optionally browser_network_requests), which the agent then analyzes to drive planning and code generation, so arbitrary third-party pages could indirectly inject instructions.

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

• Potentially malicious external URL detected (high risk: 0.80). The skill explicitly navigates at runtime to a user-supplied target_url (e.g., https://news.ycombinator.com) and extracts DOM/accessibility/screenshot data which is injected into the agent's plan/code-generation flow, meaning fetched remote page content can directly control prompts and the generated code.