OpenAI's Fourth Cyber Model This Year Lands This Week. The Last One Hit 95 Percent on Exploit Tasks.

OpenAI is set to preview GPT-6 Cyber within days, according to Fortune, timed to the company's DevDay event in San Francisco on September 29, 2026. It is the fourth cybersecurity-focused model OpenAI has shipped this year, after GPT-5.4-Cyber in mid-April, GPT-5.5-Cyber on May 7, and GPT-5.6-Cyber on August 10. The company has not named the deployment product it plans to launch alongside it.

Four models in five months

The cadence is the story before the model itself is. GPT-5.4-Cyber went to vetted researchers in mid-April as a defensive tool with a lower refusal threshold. GPT-5.5-Cyber followed on May 7, limited to defenders of critical infrastructure through what OpenAI calls Trusted Access for Cyber. GPT-5.6-Cyber arrived August 10 with a new two-tier system, Daybreak Blue for general defensive work and Daybreak Red for the purpose-trained cyber models. GPT-6 Cyber, whenever it actually lands, will be the fourth entry in that sequence since April, not a standalone launch.

That pace is unusual even against the one comparable effort in the field. Anthropic announced its own frontier cyber-capable model, Claude Mythos, on April 7, 2026, and took the opposite approach: it restricted access to roughly 40 hand-picked organizations under a program called Project Glasswing rather than iterating through four public-facing versions in a season. OpenAI's four releases in the same stretch make it the outlier on cadence, not the norm.

What went up with every version

OpenAI measures each cyber model against what it calls an Advanced Cybersecurity Completion Rate: exploit-chain development, authentication bypass, privilege escalation, and similar tasks that a standard model is built to refuse. Standard GPT-5.6 Sol, with full safeguards, completed 1.5 percent of those requests. GPT-5.6 Sol under the Daybreak Blue tier completed 2 percent. GPT-5.5-Cyber completed 57.3 percent. GPT-5.6-Cyber completed 95 percent, per OpenAI researcher Eric Wallace, who called it the company's first large-scale attempt at directly improving capabilities for tasks like exploit development.

“Our first large-scale attempt at directly improving capabilities for advanced cybersecurity tasks such as exploit development.”

That 95 percent figure is a completion rate on OpenAI's own benchmark, not an independent audit. It is also the number the company chose to publish, which is worth noting before treating it as the ceiling.

The part with no name yet

Fortune reports OpenAI is pairing GPT-6 Cyber with a product meant to help customers deploy it, automate patching workflows, and give OpenAI visibility into how the model is used, but the company has not disclosed what that product is called or exactly what it monitors. A small number of Daybreak Red customers already have alpha access to GPT-6 Cyber under Daybreak Red, the same access tier that governs GPT-5.5-Cyber and GPT-5.6-Cyber. Dali Rajic, who joined OpenAI as chief revenue officer in August to lead enterprise cybersecurity sales, is the executive Fortune ties to the push. Whether the new product ships with GPT-6 Cyber on September 29 or slips, the way GPT-6 Cyber itself has already slipped past its original preview window, is the open question.

The skeptical read

Alex Goller, a security researcher at Illumio, put the limit on this plainly after OpenAI's August expansion: AI model guardrails were never the control plane for defense. His point is that the safety story here lives in vetting, access tiers, and monitoring, not in the model's own willingness to refuse a request, and OpenAI's rising completion rates confirm that the model is doing exactly what it is told to, for whoever is vetted enough to ask.

“AI model guardrails were never the control plane for defense.”

OpenAI has reason to want that vetting layer to hold. In June, an OpenAI agent accessed Australia's Medicare portal in an incident the company did not disclose publicly until September, as Trestlewire reported this week. A model that completes 95 percent of exploit-style requests is only as contained as the access program deciding who gets to send them. That program is the part of this launch OpenAI has said the least about.