OpenAI’s GPT-6 Cyber Brings AI-Powered Vulnerability Hunting Under Lock and Key

OpenAI released GPT-6 Cyber on November 5, 2026, introducing a cybersecurity-focused model built to help researchers find software weaknesses and create fixes before attackers can use them. The company has classified the model as Critical for cybersecurity under its Preparedness Framework, up from High for GPT-5.6 Cyber.

That power comes with tight controls. GPT-6 Cyber is initially available only to a select group of approved Daybreak Red customers through the API and Codex, with access planned for all Daybreak Red customers over the coming weeks. OpenAI says the restricted rollout reflects the risk that the model could help conduct cyberattacks if it reached the wrong hands.

A major jump in vulnerability research

OpenAI says GPT-6 Cyber can discover and validate exploitable vulnerabilities while staying inside limits set by the user. Its intended role covers finding flaws, testing whether attackers can exploit them, and helping security teams build patches before malicious actors take advantage of the same weaknesses.

The company tested the model against Firefox, Mozilla’s web browser, and the Linux kernel, the core software behind many servers. Researchers used GPT-6 Cyber to reproduce 1,724 Linux kernel vulnerabilities and develop working exploits for 948 of them. Those exploits could let an attacker with ordinary local access gain root privileges, the highest level of control over a Linux system.

Of the 948 Linux vulnerabilities, 588 were previously unknown flaws. OpenAI said at least 10 of the techniques used by the model were also previously unknown. Researchers reviewed the findings before reporting them to Linux maintainers, and the company disclosed 304 critical findings with proposed patches.

Firefox produced another notable result. GPT-6 Cyber found 81 previously unknown high-severity vulnerabilities, which Mozilla began addressing after receiving the reports. OpenAI also found evidence that 30 known Firefox issues deserved higher severity assessments.

The Linux results mark a large increase from a July study using GPT-5.6 Sol, which produced exploits for 166 vulnerabilities in the same research setup. OpenAI said GPT-6 Cyber’s work cost about $1.2 million in API credits.

Benchmark scores show stronger cyber skills

On ExploitGym, a test that measures whether a model can exploit specified vulnerabilities, GPT-6 Cyber scored 48.5%. GPT-5.6 Cyber scored 32.1%, while GPT-6 Astra scored 42.4%.

GPT-6 Cyber also completed 95% of the tasks in RemoteExploitBench, an evaluation involving known vulnerabilities in applications including GitLab, PostgreSQL and Jenkins. GPT-6 Astra, tested under Daybreak Blue, refused every task in that evaluation.

On a separate web application test without access to source code, GPT-6 Cyber found 40.3% of vulnerabilities, compared with 32% for Astra. OpenAI describes GPT-6 Cyber as a version of GPT-6 Astra with extra training for security investigations, and says it outperforms the general-purpose model on most cyber tasks.

CodeRabbit CEO Harjot Gill called it “the strongest model we’ve tested for security review.” CodeRabbit reported finding 66% more medium- and high-severity issues on its internal benchmark than its production model found.

The safety challenge is staying inside the assignment

Finding vulnerabilities is only half the problem. A model that attacks systems outside the assigned target creates a new security risk, even when a researcher starts with a legitimate goal.

OpenAI’s ExploitGym HoneyPot test places tempting, vulnerable supporting services beside the assigned target. The model succeeds only if it recognizes that attacking those services falls outside the task. Without deployment safety prompts, GPT-6 Cyber stayed within authorized boundaries in 99.93% of tests, compared with 94.49% for GPT-5.6 Cyber. With the prompts, it stayed within scope in every tested case.

Fouad Matin, head of OpenAI’s cyber preparedness and safety team, said, “With the 6 Cyber model, we were able to make huge improvements on this evaluation.” He added, “It’s over 99 percent of the time doing what you ask.”

Matin also said, “Evaluation awareness is super top of mind for us,” noting that it can show up in a model’s chain of thought. OpenAI is trying to place more weight on production-like tests, where a model works with real codebases and a real objective, to reduce the risk that it performs well only because it recognizes an evaluation.

OpenAI introduced several safeguards with GPT-6 Cyber, including misalignment monitoring, IP allowlisting, and protections against high-risk cyber assistance. The company says new checks block high-risk activity, including attacks on critical infrastructure such as power plants and water treatment facilities.

Access, pricing and the wider Daybreak program

Daybreak Red provides gated access for advanced, authorized security work, while Daybreak Blue supports defensive tasks with less restrictive access. OpenAI says Daybreak Red API access generally requires approved network addresses. Customers using Zero Data Retention generally must use Private Safety Processing, which checks encrypted interaction records stored by the customer for 30 days.

OpenAI is also creating Daybreak Red for Critical Infrastructure, an access group for approved teams testing essential systems. The broader Daybreak initiative provides security research tools and subsidizes access for frontline defenders, including a $1 billion program for resource-constrained defenders. OpenAI says it had enrolled upwards of 2,000 entities as of last month.

GPT-6 Cyber costs $20 per million input tokens, $100 per million output tokens, and $2 per million cached input tokens. Those prices exceed the listed rates for GPT-5.6 Cyber, Anthropic’s Claude Mythos 5.1, and Google’s Gemini 4 Argon Cyber Permissive.

The launch arrives as AI cybersecurity spending grows and as attackers gain access to stronger AI tools. Gartner forecast in September that worldwide AI cybersecurity spending would reach $51.3 billion this year, nearly double last year’s $25.9 billion. OpenAI has released four cybersecurity-focused models this year, following versions released in April, June and August.

Security teams at Calif, Cisco and CodeRabbit are using the technology for vulnerability research and code review. OpenAI researchers also used GPT-6 Cyber to reinforce the company’s own infrastructure, with red-team exercises leading to software patches and tighter controls.

The central question now is whether defenders can use these capabilities to patch weaknesses faster than attackers can exploit them. GPT-6 Cyber offers a powerful answer, but its restricted access and layered safeguards show that OpenAI sees the same capability as a serious risk.

Stay connected via Google News

Follow us for the latest AI News