Daybreak Red is intended for more advanced authorized security work, including exploit development and red teaming. Its newest model, GPT-5.6-Cyber, is based on GPT-5.6 Sol but has received additional training for specialized tasks such as discovering previously unknown vulnerabilities and developing exploit chains. OpenAI also trained the model to refuse fewer requests involving higher-risk cybersecurity activities.
The difference is particularly pronounced in OpenAI’s internal Advanced Cybersecurity Completion Rate evaluation. GPT-5.6-Cyber completed 95% of requests involving areas such as exploit-chain development, authentication bypass and privilege escalation. GPT-5.6 Sol completed 1.5%, while GPT-5.6 Sol running through Daybreak Blue completed 2%. The previous GPT-5.5-Cyber model reached 57.3%.
That increased willingness to complete advanced requests does not translate into superior results on every security benchmark. GPT-5.6-Cyber surpassed GPT-5.6 Sol and GPT-5.5-Cyber on ExploitGym, which tests whether agents can convert known vulnerabilities into working exploits capable of arbitrary code execution in controlled environments.
OpenAI also reported stronger results for GPT-5.6-Cyber on an internal zero-day evaluation. Models were given the current release of an open-source repository and asked to identify vulnerabilities, produce proof-of-concept exploits and document their findings. The specialized model scored higher than GPT-5.6 Sol through Daybreak Blue, which OpenAI attributed to its cybersecurity-specific training.
Results were more mixed on a separate vulnerability discovery and report-writing test. Both GPT-5.6 Sol and GPT-5.6-Cyber improved on GPT-5.5-Cyber, but GPT-5.6-Cyber scored below GPT-5.6 Sol. OpenAI said the specialized model sometimes generated shorter and less detailed vulnerability reports.
GPT-5.6 Sol also retained an advantage on ExploitBench, which tests development of a V8 vulnerability into a complete exploit under more difficult conditions than ExploitGym. With agents limited to 300 turns, the Daybreak Blue version of GPT-5.6 Sol performed best and used tokens more efficiently. Increasing the limit to 600 turns reduced the performance difference between the models.
OpenAI said it has already given trusted customer partners early access to GPT-5.6-Cyber. SpecterOps CTO Jared Atkinson said the model had improved the company’s vulnerability research workflows.
“ is materially improving our specialist vulnerability-research workflows: it reasons more accurately about real exploit constraints, tracks complex state better, and has completed work in under a day that earlier models had not resolved after weeks of intermittent effort,” Atkinson said. “In a governed Trusted Access environment, reducing unnecessary refusals helps authorized researchers preserve momentum and spend more time validating findings and turning them into defensive value.”
OpenAI has also used GPT-5.6-Cyber for its own vulnerability research. The company said the model helped investigate V8, the JavaScript engine used by Chrome, and uncover two previously unknown vulnerabilities that could be combined to corrupt memory and escape the V8 heap sandbox. Researchers validated the findings and disclosed them to Google, which fixed the vulnerability identified as CVE-2026-15903.
OpenAI described CVE-2026-15903 as a high-severity V8 flaw involving an optimizing compiler that could skip a safety check while converting values to integers. Under the affected conditions, an undefined value could produce an unexpectedly large number. If used as an array index, that value could cause the compiler to omit a bounds check, potentially allowing memory belonging to other objects to be read or overwritten. OpenAI said GPT-5.6-Cyber also identified the second vulnerability needed to escape the heap sandbox.
The company said it has used the model to uncover additional security issues, including at least five vulnerabilities in a popular mobile operating system, three critical vulnerabilities in a popular database and more than 400 vulnerabilities capable of producing privilege escalation in a widely used operating-system kernel. OpenAI said it is working with Daybreak partners and open-source community members on disclosure and remediation.
Despite the model’s expanded capabilities, OpenAI said GPT-5.6-Cyber remains below the highest cybersecurity risk level in its Preparedness Framework. Both GPT-5.6 Sol and GPT-5.6-Cyber were assessed at the High capability threshold rather than Critical. OpenAI said GPT-5.6-Cyber improved on some of the specialized tasks targeted during training, but those gains were not enough to move it into the Critical category.
OpenAI also said GPT-5.6-Cyber was not involved in exploiting Hugging Face and that no other models planned for an upcoming release were involved. A more detailed system card covering GPT-5.6-Cyber evaluations is planned for a later date.
The expanded access comes with additional security requirements because Daybreak gives approved users models operating with fewer restrictions than standard deployments. OpenAI said access to both tiers is limited to approved individuals and organizations performing authorized work, with controls including identity verification, account security measures, monitoring, use restrictions and legal attestations.
Individual Daybreak accounts will be required to use hardware security keys beginning September 1, 2026. OpenAI is also encouraging customers using Codex to adopt auto-review rather than full-access mode. Auto-review checks actions that require elevated permissions before they run and can stop requests deemed to carry significant destructive risk.
OpenAI said further monitoring measures are being developed and that additional security improvements are expected in the coming weeks. It is also prioritizing alignment training and testing for future Daybreak releases and recommends that security teams isolate model-driven workflows, monitor agent actions and clearly define which systems and activities are authorized.
For most defensive security work, OpenAI recommends starting with Daybreak Blue. Organizations conducting authorized advanced vulnerability research, exploit development or red-team exercises can request Daybreak Red, giving them access to GPT-5.6-Cyber and the company’s most capable specialized cybersecurity models.
This analysis is based on reporting from OpenAI.
Image courtesy of OpenAI.
This article was generated with AI assistance and reviewed for accuracy and quality.
