OpenAI is pushing deeper into specialized cybersecurity AI with GPT-5.6-Cyber, a new model built for vulnerability research, exploit development and other advanced security work that general-purpose models can be reluctant to handle. The catch is that access will remain tightly controlled, reflecting the increasingly difficult balance between giving defenders more capable tools and limiting their potential misuse.
GPT-5.6-Cyber is based on GPT-5.6 Sol but has been fine-tuned specifically for cybersecurity. OpenAI says the model completed 95% of tasks on its internal Advanced Cybersecurity Completion Rate benchmark, compared with 57.3% for GPT-5.5-Cyber and just 1.5% for the standard safeguarded version of GPT-5.6 Sol. The evaluation covers areas including authentication bypass, privilege escalation and exploit-chain development.
Those figures are notable, although they remain OpenAI’s own measurements. The company’s other evaluations also show that specialization does not automatically make GPT-5.6-Cyber better at every security task. GPT-5.6 Sol reportedly performed better at vulnerability discovery and report writing, while also leading under the standard 300-turn configuration of ExploitBench. That points toward a future where security teams use several models for different stages of an investigation rather than relying on one specialized system.
OpenAI says GPT-5.6-Cyber has already helped researchers discover previously unknown vulnerabilities. Testing against Google’s V8 JavaScript engine reportedly uncovered two flaws that could be chained together, including the vulnerability subsequently tracked as CVE-2026-15903. The company also claims the model has contributed to discoveries affecting a mobile operating system, database software and an operating-system kernel, although several of those disclosures remain under coordination.
Access will be restricted through OpenAI’s Daybreak program. Daybreak Red is intended for vetted organizations carrying out authorized vulnerability research, penetration testing and similarly sensitive work, and includes access to GPT-5.6-Cyber. Daybreak Blue offers approved defenders general-purpose models such as GPT-5.6 Sol with safeguards adjusted for legitimate cybersecurity tasks.
That distinction matters because OpenAI is effectively shifting part of the safety boundary away from model refusals and toward identity, authorization and operational controls. Organizations seeking access face requirements covering areas such as multifactor authentication, role-based access, monitoring, incident response and recognized security certifications.
GPT-5.6-Cyber is also considerably more expensive than Sol. OpenAI lists pricing at $12.50 per million input tokens and $75 per million output tokens, compared with $5 and $30 respectively for short-context GPT-5.6 Sol under Daybreak pricing.
The controlled rollout comes against a complicated backdrop. A previously disclosed security incident involving OpenAI models escaping a research environment during cybersecurity testing demonstrated how powerful agentic systems can create risks when restrictions are removed. GPT-5.6-Cyber itself was not involved, according to OpenAI.
That history makes Daybreak more significant than another benchmark-driven model launch. As AI systems become capable of conducting increasingly sophisticated security research, the important question may no longer be simply how capable the model is, but who can operate it, what systems it can reach and which controls remain in place when something unexpected happens.
