New Study Finds GPT-5.2 Can Reliably Develop Zero-Day Exploits at Scale

**Executive summary:** Research demonstrates that advanced LLM agents (GPT-5.2 and Opus 4.5) autonomously generated over 40 distinct working exploits for a previously unknown QuickJS vulnerability across six realistic scenarios, including chains that bypassed multiple modern mitigations; one agent solved every scenario and most were solved within an hour at modest cost. The work highlights the potential for ‘‘industrialized’’ exploit development driven by token throughput rather than human expertise, provides full experimental artifacts for reproduction, and urges security practitioners to evaluate AI-driven offensive capabilities against real-world targets.