OpenAI Expands Daybreak Cybersecurity Programme, Launches More Permissive GPT-5.6-Cyber Model

OpenAI has expanded its Daybreak cybersecurity initiative and introduced GPT-5.6-Cyber, a specialised version of its GPT-5.6 Sol model trained to reduce refusals on advanced, dual-use security tasks such as exploit-chain development and vulnerability research, the company announced in a blog post dated August 10, 2026. The move is aimed at giving vetted defenders frontier AI capabilities before attackers can deploy similar tools at scale, OpenAI said.

Two access tiers for vetted defenders

Daybreak now offers two tiers. Daybreak Blue gives approved users access to general-purpose models, including GPT-5.6 Sol, with safeguards tailored for authorised defensive work such as vulnerability discovery, secure code review, malware analysis and incident response. Daybreak Red provides access to OpenAI's purpose-built cybersecurity models, including the newly introduced GPT-5.6-Cyber, intended for authorised vulnerability research, exploit validation and security testing.

Sharp drop in refusals

According to OpenAI's internal Advanced Cybersecurity Completion Rate evaluation, GPT-5.6-Cyber completed 95 percent of requests involving exploit-chain development, authentication bypass and privilege escalation, compared with 1.5 percent for the standard GPT-5.6 Sol model and 2 percent under Daybreak Blue access. The company said this also marks an improvement over its predecessor, GPT-5.5-Cyber, which completed just 57.3 percent of such requests.

Real-world vulnerabilities uncovered

OpenAI said it has used GPT-5.6-Cyber to study widely used software, uncovering two previously unknown vulnerabilities in V8, the JavaScript engine that powers Chrome, which could be chained to corrupt memory and escape the browser's sandbox. The findings were reported to Google through coordinated disclosure and fixed, with the flaw assigned CVE-2026-15903. The company said it has also used the model to identify vulnerabilities in a popular mobile operating system, a widely used database, and an operating system kernel, and is working with partners to disclose and fix these separately.

Capability threshold and safeguards

Under OpenAI's Preparedness Framework, both GPT-5.6 Sol and GPT-5.6-Cyber were assessed as reaching the 'High' cybersecurity capability threshold but remaining below 'Critical'. OpenAI said GPT-5.6-Cyber was not involved in a prior Hugging Face security incident. Access to Daybreak Blue and Red is restricted to vetted individuals and organisations through identity verification and legal attestations, and the company said it will require hardware security keys for all individual Daybreak accounts starting September 1.

OpenAI cited early positive feedback from partners including SpecterOps, SentinelOne and Palo Alto Networks, with SpecterOps chief technology officer Jared Atkinson saying the model has resolved research tasks in under a day that earlier models had failed to complete after weeks of effort. OpenAI said a fuller system card evaluating GPT-5.6-Cyber will be published at a later date.