OpenAI drops GPT-6 Cyber after auto-patching 948 Linux kernel bugs

Top-tier vulnerability research usually means weeks of human experts squinting at low-level source code. OpenAI just showed what happens when you turn that work into a high-speed assembly line.

OpenAI officially introduced GPT-6 Cyber, a specialized model trained specifically for cybersecurity research, exploit validation, and patch generation. To prove its point, researchers plugged it into an autonomous vulnerability pipeline and pointed it straight at the Linux kernel.

In a single automated run, the pipeline surfaced 2,308 candidate issues, dynamically reproduced 1,724 vulnerabilities, and built 1,169 working local privilege escalation exploits. After deduplication, it confirmed 948 distinct local privilege escalation bugs in the Linux kernel.

The entire run burned through $1.2 million in API credits. When OpenAI sampled 166 of the generated exploits, they found at least 10 techniques that had never been documented before. For comparison, GPT-5.6 Sol found 166 vulnerabilities in the same harness, meaning the new model delivered a 5.7-fold increase in output.

Lethal in the sandbox, disciplined in the wild

Linux was not the only target. GPT-6 Cyber also uncovered 81 previously unknown high-severity vulnerabilities in Firefox, while demonstrating that 30 existing bugs were far more dangerous than initially assumed. Mozilla has already begun rolling out patches based on the model's reports.

On the ExploitGym benchmark, which measures remote exploitation capabilities, GPT-6 Cyber scored 48.5%, compared to 32.1% for GPT-5.6 Cyber. In black-box web penetration tests, it reached a 40.3% vulnerability recall rate. Because of these offensive capabilities, OpenAI bumped the model's cybersecurity classification under its Preparedness Framework to "Critical," up from "High."

Yet the hardest problem with autonomous security agents is restraint. In pentesting environments, rogue agents often wander off-target and attack neighboring infrastructure.

OpenAI spent significant compute aligning the model to follow engagement boundaries. In an isolated honeypot evaluation with easily hackable adjacent services, GPT-6 Cyber avoided out-of-scope targets in 99.93% of runs without deployment safety prompts, and hit a perfect 100% compliance rate once those safety prompts were applied.

Access is tightly controlled. GPT-6 Cyber is not hitting the general API; it is currently restricted to vetted Daybreak Red customers, including teams at Cisco, Palo Alto Networks, and CodeRabbit. OpenAI plans to host a dedicated cyber livestream on November 12 to dive deeper into the technical setup.

Why it matters

This marks the point where software vulnerability research crosses from artisanal craftsmanship into heavy industrial automation.

For developers and security professionals, the economics are jarring. Spending $1.2 million to harvest nearly a thousand confirmed kernel vulnerabilities translates to roughly $1,200 per zero-day exploit, far below traditional human bug bounties or black-market broker fees. But this asymmetry introduces an immediate problem: finding exploits is now automated, while validating and merging fixes in open-source projects remains stubbornly manual.

Strategically, OpenAI's playbook makes sense, even if it highlights a narrow defensive window. The company knows malicious actors will eventually distill or replicate these offensive workflows. Restricting the model to verified defenders under Daybreak Red, backed by a $1 billion grant program for critical infrastructure, is an attempt to patch core codebases before adversaries catch up.

In our view, the real technical breakthrough here is boundary adherence rather than raw exploit generation. Enterprise red teams cannot deploy an autonomous model if there is any chance it goes rogue and brings down live production servers. By pushing out-of-scope violations close to zero, OpenAI made an autonomous cyber agent that serious security teams can actually contemplate letting off the leash.