GPT-5.6-Cyber: OpenAI's Defense-Only Model Arrives as the Cyber Defense Window Narrows
On August 10, 2026, OpenAI announced GPT-5.6-Cyber, a model built only for cyber defense, along with wider access tiers for its Daybreak cybersecurity program. The official blog post is titled "Expanding Daybreak as the Cyber Defense Window Narrows" (OpenAI announcement). The X post from Tibo (@thsottiaux), who leads OpenAI Codex, passed 570,000 views (original X post), and co-founder Greg Brockman (@gdb) shared the same news (@gdb).
95.0% vs. 1.5%: Why Cyber Task Completion Rates Split So Sharply
On Advanced Cybersecurity Completion Rate, an internal metric OpenAI published, GPT-5.6-Cyber scored 95.0%. The general-purpose GPT-5.6 Sol, which it is built on, scored 1.5%. Sol accessed through Daybreak Blue scored 2.0%, and the previous GPT-5.5-Cyber scored 57.3% (OpenAI announcement).
| Model | Advanced Cybersecurity Completion Rate |
|---|---|
| GPT-5.6-Cyber | 95.0% |
| GPT-5.6 Sol (general) | 1.5% |
| GPT-5.6 Sol (via Daybreak Blue) | 2.0% |
| GPT-5.5-Cyber (previous model) | 57.3% |
These figures come from OpenAI's own internal metric, not an independent outside evaluation. Still, the model beats its predecessor's 57.3% and sits more than 60 times above the general model, which supports the view that a defense-specialized model makes a real difference on cyber tasks (SecurityWeek).
Blue and Red: Two Access Paths for Defenders
The core of this expansion is that Daybreak is now split in two. Daybreak Blue gives access to general frontier models for defensive work such as vulnerability discovery, secure code review, malware analysis, and incident response, and it is the recommended starting point for most defenders. Daybreak Red gives access to specialized models like GPT-5.6-Cyber, and it comes with strict vetting that requires identity verification, monitoring, approved use cases, and legal checks (CyberScoop). Access is broad for defense and narrow for high-risk specialized tools, so the width of access varies with the use.
A Lower-Refusal, High Dual-Use Model: What Did It Find in Practice?
GPT-5.6-Cyber is built on GPT-5.6 Sol and specializes in zero-day discovery, exploit chain development, vulnerability research, and red teaming. It is designed to refuse less often on high dual-use tasks (Engadget). The real-world examples OpenAI shared are specific. It found two previously unknown vulnerabilities in Chrome's V8 engine that, chained together, could lead to memory corruption and a sandbox escape. After they were disclosed to Google, a High-severity patch shipped under CVE-2026-15903. It also found more than five mobile OS vulnerabilities, three critical vulnerabilities in a widely used database, and more than 400 privilege-escalation issues in a popular OS kernel. The affected products were not named, and disclosure is still underway (SecurityWeek). Jared Atkinson, CTO of SpecterOps and an early customer, said, "It meaningfully improves professional vulnerability research workflows, and work that used to take weeks can be finished within a day" (CyberScoop).
Holding Back Astra but Releasing Cyber: Different Risk Assessments
The announcement came three days after OpenAI delayed the release of its Astra model over concerns about "Critical" cyber risk, meaning automated zero-day construction and end-to-end attacks. GPT-5.6-Cyber, by contrast, was rated under the Preparedness Framework as reaching the "High" cyber capability threshold alongside GPT-5.6 Sol, below "Critical," and it was stated to have had no part in the earlier Hugging Face intrusion incident (CyberScoop). In other words, the choice blocks the possibility of automated attacks while unlocking defensive capability. Controls are tightening too. Starting September 1, hardware security keys will be mandatory for all individual Daybreak accounts, and the roster of partner security firms grows to 16, including Accenture, IBM, CrowdStrike, Cisco, Palo Alto Networks, Sophos, Cloudflare, Capgemini, EY, KPMG, PwC, Akamai, Fortinet, and SentinelOne (OpenAI announcement).
Can Anyone Use Daybreak?
For defenders in practice, this announcement shows that vulnerability discovery speed and coverage have moved up a notch at the model level. Some things still need checking, though. The 95.0% figure is OpenAI's internal benchmark, and it is too early to read it as real-world performance without outside validation. It also remains to be seen how far a lower-refusal, high dual-use model will be allowed to go in production, and whether cases like the V8 finding lead to immediate disclosure and patching. Whether access opened for defense can stay out of attackers' hands ultimately depends on Daybreak Red's vetting criteria and operational controls such as hardware keys (The Next Web). If your organization uses AI for cybersecurity, this announcement is a good moment to decide which door to open, Blue or Red, and what uses to allow behind it.