Unintended real-world intrusions during Gemini evaluations expose critical blind spots in third-party AI safety protocols.