OpenAI Unveils GPT-6 Cyber to Automate Vulnerability Discovery

OpenAI has introduced GPT-6 Cyber, a specialised artificial intelligence model designed to assist security researchers in identifying, validating, and patching exploitable vulnerabilities. During initial testing, researchers used the model to uncover 81 high-severity vulnerabilities in Firefox and 948 vulnerabilities within the Linux kernel, including 588 previously unknown flaws.

The development aims to enable security operations to move at “agent speed.” According to OpenAI researcher Eric Wallace, the goal is to validate the potential exploitability of vulnerabilities at scale and prioritise the most consequential flaws. This includes testing fixes against alternative routes of exploitation to ensure that a single vulnerability does not provide multiple attack paths.

Discovery of High-Severity Vulnerabilities

Beyond the volume of bugs, the model demonstrated a capacity for advanced technical discovery. In a sample of 166 exploits, OpenAI researchers identified 10 exploitation techniques that appeared novel to external kernel exploitation experts. These techniques could potentially allow an attacker with local access to obtain root privileges.

Rob T. Lee, the chief AI officer at industrial cybersecurity firm Dragos, suggested that these findings indicate AI models may be capable of creating entirely new exploitation paths. Such developments could introduce new Tactics, Techniques, and Procedures (TTPs), which would make threat detection significantly more complex for defenders.

The Remediation Gap and Security Risks

While the model assists in finding flaws, experts warn that the speed of discovery may outpace the ability of organisations to fix them. Chris Hughes, VP of security strategy at Noma Security, noted that the findings in Firefox suggest a deeper reservoir of vulnerabilities in mature software than previously assumed.

Hughes highlighted a growing “vulnpocalypse” risk, noting that the OpenAI Patch the Planet dashboard currently shows a remediation rate of less than 25%. He cautioned that improved discovery capabilities could exacerbate this gap, meaning more vulnerabilities are found by AI but remain unfixed, leaving them vulnerable to malicious actors using similar technology.

To mitigate risks during the disclosure process, OpenAI has engaged in advance coordination with maintainers and security vendors. For example, the company worked with Cloudflare to deploy Web Application Firewall (WAF) rules for PHP vulnerabilities, providing proactive protection to websites before public patches were available.

Restricted Access and Safety Safeguards

OpenAI classifies GPT-6 Cyber as “Critical” under its own Preparedness Framework. To manage the risks associated with such powerful capabilities, the company has implemented stricter alignment and safeguards. In its ExploitGym Honeypot evaluation, the model went beyond its intended scope in only 0.07% of cases, compared to a 5.51% out-of-scope rate for the previous GPT-5.6 Cyber model.

Access to GPT-6 Cyber is currently restricted to a select group of approved Daybreak Red customers through the API and Codex. OpenAI plans to expand access to all Daybreak Red customers over the coming weeks and is also making GPT-5.6 Cyber available to this customer set. The company is also broadening alpha access to Daybreak Blue for Astra, its general-purpose GPT-6 frontier model.