OpenAI's GPT-6 Cyber Reproduced 1,724 Kernel Bugs as Agent Tools Launch

OpenAI says GPT-6 Cyber, its restricted vulnerability-research model, reproduced 1,724 Linux kernel vulnerabilities and built working exploits for 948 of them — a capability figure that now sits behind two commercial security agents the company launched on Thursday, Nov. 12, at its "Cyber at the Frontier" event.

The two products extend Daybreak, the cyber-defense program OpenAI launched earlier this year. Daybreak Red Team runs AI agents that penetration-test web applications. Daybreak Defense Factory investigates security findings, drafts patches and opens pull requests. Together they connect finding a vulnerability with fixing it, while engineers approve every code change.

The launch follows the release of GPT-6 Cyber by one week. OpenAI rates the model "Critical" for cybersecurity under its Preparedness Framework, the highest risk tier, and says it has helped uncover thousands of vulnerabilities in widely used software, including Linux and Firefox.

What does Red Team actually do?

Daybreak Red Team sends multiple AI agents after a web application. "Agents use source code and app context to investigate attack paths and confirm vulnerabilities," OpenAI says. The product ingests findings from code scanners, bug bounty submissions and vulnerability disclosure reports, then tests whether each one is an exploitable risk or a false alarm — a task that consumes much of a security team's time.

It can also retest the patches meant to close those holes. The agents run in isolated environments, and OpenAI says a "Guardian" review layer enforces the testing boundaries and controls which applications and endpoints the agents may touch. That safeguard keeps an automated attacker inside the scope a customer has authorized. OpenAI is accepting requests for early access.

How does Defense Factory turn findings into fixes?

Defense Factory handles the repair side. "Agents trace vulnerabilities through your code and develop patches," OpenAI says. It works on findings from Red Team and from other security tools, importing vulnerability data from Wiz, Snyk and Endor Labs. It can also scan connected GitHub repositories, watch new commits and run scheduled security assessments.

Engineers review the evidence, edit a patch with OpenAI's Codex coding agent if needed, and receive the fix as a draft GitHub pull request. Nancy Wang, who leads cyber-applied AI at OpenAI, said engineers get a merge-ready pull request with the patch and a test showing that the fix works. Patches are never deployed automatically.

Defense Factory ships as a preview plugin inside ChatGPT for workspaces that already have Daybreak access. OpenAI has not published pricing for either product, and the coverage names no launch customers.

How strong is the model underneath?

In a mid-October research run on the Linux kernel, OpenAI says GPT-6 Cyber flagged 2,308 potential flaws, reproduced 1,724 vulnerabilities and developed local privilege escalation exploits for 948 of them. That is more than five times the 166 exploits managed by GPT-5.6 Sol with the same research harness.

The model also found 81 previously unknown Firefox bugs that Mozilla rated high severity, and it exploited 48.5% of vulnerabilities on the ExploitGym benchmark, against 32.1% for GPT-5.6 Cyber. All of those figures come from OpenAI's own testing.

Access to GPT-6 Cyber itself remains limited to approved customers in Daybreak Red, OpenAI's tier for authorized exploit development and penetration testing. Expansion to all Daybreak Red customers is promised but undated. "It is clear that we have a limited window to use frontier models to secure software," OpenAI said when it released the model.

Who else is moving in this space?

Rival labs are making the same pitch to defenders. Anthropic and Google researchers found the critical HTTP/2 flaw that Apache patched on Nov. 12, and OpenAI signed JPMorgan Chase for Codex and its cyber models earlier this month.

The offensive side is moving as well. Benchmarking firm Vals found a free DeepSeek model came within eight points of OpenAI's restricted cyber model on network hacking. CloudSEK said Kazakhstan-linked hackers ran an AI agent on DeepSeek against Central Asian governments. On the same day as OpenAI's event, Trump's AI task force met to discuss AI-powered cyberattacks, Reuters reported.

With pricing unpublished and no named customers, OpenAI has yet to show commercial traction for the agent pair — but the kernel exploit figures and the JPMorgan signing signal that AI-driven security operations have moved from research demos to product roadmap.