GPT-6 Cyber Explained: OpenAI's New Bug-Hunting Model

OpenAI released GPT-6 Cyber on Thursday, a security-tuned frontier model that the company says is better at finding and proving serious vulnerabilities, and better at staying inside the limits a security team sets for it. The number getting the most attention is 948, the count of Linux kernel flaws OpenAI's automated pipeline turned into working exploits, then handed to reviewers with candidate patches. For defenders, the story is less about one model than about how cheap and repeatable large-scale bug hunting has become.

• GPT-6 Cyber scored 48.5% on OpenAI's ExploitGym evaluation, up from 32.1% for GPT-5.6 Cyber, according to The Register
• OpenAI says its researchers used the model to find 81 previously unknown high-severity Firefox vulnerabilities, and Mozilla has started fixing them
• A mid-October Linux kernel run validated 948 local privilege escalation bugs for $1,213,430 in API credits, about $1,279 per exploit
• Access starts with a select group of approved Daybreak Red customers through the API and Codex, then expands to all Daybreak Red customers over the coming weeks

What Is GPT-6 Cyber?

GPT-6 Cyber is OpenAI's latest model for authorized security work such as exploit validation, penetration testing and red teaming. OpenAI rates it "Critical" under its Preparedness Framework, the top risk tier for cyber capability, because the model can identify and develop functional zero-day exploits. That rating explains the careful rollout through OpenAI's Daybreak Red tier, which is limited to organizations cleared for offensive security testing. GPT-5.6 Cyber is also becoming available to that group.

OpenAI trained the new model to stay within its authorized scope and report results accurately. In a separate evaluation, it avoided exploiting out-of-scope services more than 99% of the time, and 0.07% of cases went beyond the exercise's intended scope. The Register notes that OpenAI declined to give further detail on those cases, so that figure is best read as the company's own measurement rather than an independent audit.

How the Linux Kernel Patch Pipeline Works

OpenAI's research write-up describes a chain of specialized AI agents. They narrow the kernel down to security-relevant code, scan it for candidate bugs, reproduce those bugs in an isolated environment, assess whether they can be exploited, and draft a fix. A human security researcher then reviews the output before anything goes to maintainers.

In the October run, the pipeline flagged 2,308 candidates and reproduced 1,724 of them. It also produced proof-of-concept code for 291 information leaks. OpenAI says the agents discovered at least 10 novel exploitation techniques, with the technical details withheld for safety. For comparison, a late-July scan using GPT-5.6 Sol yielded 166 unique exploitable bugs at a cost of $513,690, according to SecurityWeek.

Following guidance from the kernel community, OpenAI started with critical flaws and says it has already submitted hundreds of reports and patches, with the rest following in stages alongside its Patch the Planet partners. The Register reports that 588 of the 948 bugs were previously unknown. The work complements the company's recent bug bounty increase, which rewards outside researchers for the same kind of discovery.

What Should Security Teams Do Now?

The practical takeaway is that patch speed matters more than ever. OpenAI advises treating a low-privileged account on an unpatched operating system as functionally equivalent to root access for a well-resourced attacker. It points to layered defenses, including sandboxing and isolation tools such as Firecracker and gVisor, to contain the damage when a flaw is exploited.

That makes three habits worth prioritizing: keep Linux hosts on current kernels, reduce the number of low-privileged accounts that can reach sensitive systems, and isolate untrusted workloads. For more on how defenders are adopting tools like these, see our AI security coverage.

Sources: The Register — November 5, 2026; SecurityWeek — November 5, 2026; OpenAI — November 4, 2026.