OpenAI Launches GPT-6 Cyber, Its Most Powerful Security Model Yet

OpenAI has launched GPT-6 Cyber, a specialised artificial intelligence model built to hunt for software vulnerabilities, prove they can be exploited, and help fix them before attackers strike. The company announced the model on 5 November, a day after publishing the research behind it, and immediately gave it the highest cybersecurity rating in its own safety framework.

GPT-6 Cyber is a cyber-focused version of GPT-6 Astra, OpenAI’s main model, trained for authorised vulnerability research. Unlike a general-purpose chatbot, it is designed to work inside limits set by security researchers: it searches code for flaws, builds working exploits to confirm them, and proposes patches.

“We are focused on creating a thousand fewer paths to root for attackers while empowering defenders,” OpenAI said in its announcement.

948 working exploits from the Linux kernel

The headline evidence for the model comes from a controlled research campaign described in a companion post, “A thousand fewer paths to root”, published on 4 November. In that campaign, an automated pipeline built on GPT-6 Cyber scanned about 2.8 million lines of Linux kernel code — roughly 7 per cent of the project — and turned 2,308 candidate findings into 1,724 reproduced vulnerabilities. From those, it developed working exploits for 948 distinct local privilege escalation flaws, attacks that let an ordinary local user gain root access.

OpenAI says 588 of the 948 flaws were previously unknown, and it disclosed 304 critical findings to Linux maintainers along with candidate patches. The whole run cost $1,213,430 in API credits, or about $1,279 per working exploit. The company notes that figure is more an upper bound than a precise price, since the pipeline was built for research rather than efficiency.

The scale is a sharp jump from the previous generation. A July run of the same test harness with GPT-5.6 Sol found and exploited 166 bugs for $513,690.

OpenAI’s own conclusion is blunt: in the near term, a low-privileged account on a system that is behind on patches should be treated as equal to root for a well-funded attacker. The figures are the company’s own claims and have not been independently audited.

81 new flaws found in Firefox

In a separate campaign, GPT-6 Cyber produced evidence for 81 previously unknown vulnerabilities in Firefox that Mozilla classified as high severity, along with proposed patches. Mozilla has since shipped fixes for many of them. The company’s advisory MFSA 2026-118 covers fixes in Firefox 159, released on 27 October.

On the ExploitGym benchmark, which tests whether a model can exploit a specified vulnerability, GPT-6 Cyber scored 48.5 per cent, up from 32.1 per cent for GPT-5.6 Cyber — a gap OpenAI calls about 16 percentage points and a 51 per cent relative improvement. On RemoteExploitBench, a test of remote exploitation, it completed 95 per cent of tasks.

The launch lands as the security industry watches AI’s growing role on both sides of the attack. It follows a wave of AI security deals, including CrowdStrike’s near-$500 million bet on securing AI agents, and comes as investigators continue to chase the record $6.3 billion Binance theft, with the FBI publishing hundreds of wallet addresses linked to the heist.

A “Critical” rating — and tightly limited access

OpenAI has classified GPT-6 Cyber as “Critical” for cybersecurity under its Preparedness Framework, its highest designation, up from “High” for GPT-5.6 Cyber. The rating is the company’s own assessment, not an outside certification, and it is the reason access is so tightly controlled.

GPT-6 Cyber is initially available through the API and Codex, OpenAI’s coding agent, to a select group of approved Daybreak Red customers, with identity checks, organisational security requirements and API IP allowlisting. Daybreak Red is the tier of OpenAI’s Daybreak programme for offensive-style authorised work such as penetration testing and red teaming; a separate Daybreak Blue tier serves verified defenders under more restrictive rules. OpenAI says access will reach all Daybreak Red customers over the coming weeks.

The company says it has added safeguards for industrial controls, essential services, election systems and other high-risk targets. There is no general ChatGPT version of the model and no public price. Neowin reported it is not being offered as an unrestricted public tool, while Cybersecurity News reported access begins only with vetted organisations.

What it means for defenders

For security teams, the model promises to compress weeks of expert labour into hours — finding flaws, proving they matter, and drafting fixes. For everyone else, the message is less comfortable: the same capability in the wrong hands would be a powerful weapon, which is why OpenAI is keeping it behind approval checks.

The launch comes as OpenAI’s data-centre expansion continues.

Whether GPT-6 Cyber shifts the balance toward defenders will depend on how quickly its findings turn into patches — and on whether the guardrails around it hold.

For now, the model demonstrates what frontier AI can already do to the software the world runs on — and warns that unpatched systems are running out of time.