What 20 years of kernel bugs taught us about finding the next one | Jenny Qu | Bug Bash 2026

Security researcher Jenny Qu analyzes how large language model integrated systems are repeating the same fundamental security mistakes that plagued the Linux kernel for twenty years. She analyzed 125,000 kernel bug fixes and found that 13% of bugs remained hidden for over five years, with some persisting for 19 years. These long-hidden bugs share a common pattern: data and control paths collapse into one channel, with only programmer intent distinguishing between them. Qu demonstrates how this exact vulnerability class now exists in every LLM-integrated system, where natural language input can simultaneously carry data and control instructions through tokenization, showing a Microsoft Copilot vulnerability (CVE-2025-32711) that achieved zero-click exploitation through prompt injection via email content. She argues that current defenses focusing on content classification are fighting on the wrong axis, and proposes three solutions: building proper test harnesses for hostile content injection, defining privilege boundaries for LLM actions, and implementing capability-based authority with separated control planes.