OpenAI Unveils GPT-5.6-Cyber as It Uncovers 400+ Kernel Vulnerabilities
While engineers were busy patching minor bugs, OpenAI quietly unleashed an AI model dedicated to offensive hacking. The results are both terrifyingly impressive and a total wake-up call for cyber defense.
As part of its new Daybreak cybersecurity initiative, OpenAI split its access tiers into two distinct flavors: Daybreak Blue running GPT-5.6 Sol for routine code auditing, and Daybreak Red powered by GPT-5.6-Cyber for high-stakes exploit research.
In internal benchmarks measuring advanced cyber capabilities like authentication bypass and privilege escalation, the new GPT-5.6-Cyber scored a staggering 95% completion rate, utterly crushing the older GPT-5.5-Cyber at 57.3% and leaving GPT-5.6 Sol in the dust at under 2%.
To prove its real-world capability, the AI was let loose on Chrome’s V8 engine, where it promptly discovered two zero-day memory corruption flaws that could break out of the heap sandbox, prompting Google to issue an immediate patch for CVE-2026-15903.
Without revealing affected vendors just yet, OpenAI confirmed the model also unearthed five vulnerabilities in a major mobile OS, three critical database flaws, and over 400 privilege escalation bugs in a widely used operating system kernel.
Human developers have spent decades trying to secure the foundational software running the modern world, only for an artificial brain to tear it apart in an afternoon. The race between automated exploit generation and defensive patching just skipped a few gears, leaving software security hanging by a very thin thread.
Source: OpenAI
Comments
This is where the magic happens: AI reads your discussion and rewrites the article based on the most interesting comments. Each strong comment adds points to the meter below. Once the meter is full, the article updates live — no page reload needed.