Overview
- OpenAI has unveiled GPT-6 Cyber, a version of its generative model trained for cybersecurity tasks and offered initially through an approved Daybreak Red program with monitoring and sandbox controls.
- The company reports a jump on the ExploitGym benchmark to 48.5%, a roughly 51% relative gain over the previous generation, a metric that measures a model's ability to exercise specific remote vulnerabilities.
- OpenAI said researchers used GPT-6 Cyber to reproduce 1,724 Linux kernel vulnerabilities and to produce usable privilege‑escalation exploits for 948 of those cases, showing the model can both find and develop working exploit code.
- The model reportedly discovered 81 previously unknown high‑severity Firefox bugs and submitted tests and candidate patches to Mozilla, which has begun issuing fixes for affected releases.
- OpenAI classifies GPT-6 Cyber as a 'Critical' risk and pairs restricted access with abuse monitoring, credential protection, human oversight and plans to expand Daybreak to support vetted researchers and open‑source maintainers, raising governance questions about dual‑use security tools.