OpenAI’s research shows systematic vulnerability discovery can be scaled, moving beyond per‑finding manual investigations. The team built a prototype to scan the Linux kernel in a controlled environment, identifying 2,308 candidate findings and reproducing 1,724 of them dynamically.
The prototype produced 948 local privilege escalation (LPE) exploits that could let an unprivileged local user gain root access. Using this evidence, OpenAI prioritized the most critical flaws for reporting and generated candidate patches for maintainer review. The effort follows earlier Patch the Planet work where GPT‑5.5 Cyber uncovered 24 LPE exploits and expanded into a full pipeline.
The Linux kernel was chosen because it powers Android devices, cloud infrastructure, supercomputers, and billions of systems. With over 40 million lines of code, its attack surface includes system calls, networking, filesystems, and device drivers. Frontier models demonstrated they can locate vulnerabilities at this scale, but the bottleneck remains validation and prioritization. OpenAI’s pipeline consists of five stages—discovery, dynamic validation, exploitability assessment, patch generation, and human review—each with cost metrics to help defenders assess resources. The goal is to reduce attacker opportunities and raise exploitation costs.
Source: OpenAI Blog · Summarized by HeadlinesBriefing