OpenAI представила модель GPT-5.6-Cyber, ориентированную на кибербезопасность

OpenAI has introduced GPT-5.6-Cyber, a specialized model designed to assist authorized security teams with vulnerability research and incident response. Built on the GPT-5.6 Sol foundation, this variant reduces refusals for dual-use tasks such as exploit development, achieving a 95% completion rate on internal cybersecurity tests compared to 1.5% for its general-purpose counterpart.

During testing, the model identified CVE-2026-15903, a critical out-of-bounds read/write flaw in Google's V8 engine that enables remote code execution within the sandbox. While standard models like GPT-5.6 Sol still perform better in open-ended repository analysis, GPT-5.6-Cyber is currently restricted to trusted enterprise partners under the 'Daybreak Red' access tier, aiming to close the gap between discovery and patching before attackers can act.