OpenAI makes GPT-5.6-cyber less restricted for trusted cybersecurity researchers: All details
OpenAI's GPT-5.6-Cyber is built on GPT-5.6 Sol and specifically trained for cybersecurity-related tasks, such as finding zero-day vulnerabilities and developing exploit chains.
- Aug 11, 2026,
- Updated Aug 11, 2026 3:26 PM IST
OpenAI has expanded its Daybreak Program with a new specialised cybersecurity model, GPT-5.6-Cyber, and two new access tiers built for cybersecurity professionals. This will give cyber defenders access to more capable tools before attackers can weaponise similar AI systems at scale.
OpenAI Daybreak: Two Levels of Access
OpenAI’s Daybreak Program now offers two tiers of access: Daybreak Blue and Red. The Blue tier gives security researchers access to general-purpose advanced AI models, including GPT-5.6 Sol, to ensure safeguards such as finding software vulnerabilities, reviewing code, analysing malware, and responding to cybersecurity incidents.
Must read: OpenAI brings Premium Seats to ChatGPT Business with 5x more AI usage
The Daybreak Red offers greater specialised access, offering purpose-built cybersecurity models such as the new GPT-5.6-Cyber. This model can be used for vulnerability research, testing whether exploits actually work, and security testing.
GPT-5.6 Cyber: Fewer refusals, sharper exploit skills
The GPT-5.6-Cyber is built on GPT-5.6 Sol and is specifically trained for cybersecurity, such as finding zero-day vulnerabilities and developing exploit chains, but with fewer restrictions. During internal evaluations, OpenAI found that GPT-5.6-Cyber completed 95% of advanced cybersecurity requests. Whereas, safeguarded GPT-5.6 Sol completed only 1.5%. In addition, its predecessor, GPT-5.5-Cyber, completed around 57%.
Researchers are also using GPT-5.6 Cyber for vulnerability research. They also found two previously unknown vulnerabilities in V8, the JavaScript engine used by Chrome.
The vulnerabilities could reportedly be chained together to corrupt memory and escape Chrome's security sandbox. The vulnerabilities were reported to Google and subsequently fixed.
Must read: OpenAI rolls out GPT-5.6 Luna as default ChatGPT model, removes text chat limits for free tier users
It also helped find vulnerabilities across major mobile operating systems, a popular database, and an operating-system kernel, including hundreds of privilege-escalation vulnerabilities.
Unlike the standard GPT-5.6 Sol, which refuses sensitive cybersecurity requests due to fear of misuse, GPT-5.6-Cyber is designed to refuse fewer legitimate dual-use security requests. While OpenAI is reducing some model restrictions, it is also introducing added safeguards, such as hardware security keys, auto-review in Codex, and it also plans to expand monitoring of how the model is being used.
However, access to GPT-5.6 Cyber will remain limited to cybersecurity individuals and organisations with identity verification and signed legal attestations confirming authorised use.
For Unparalleled coverage of India's Businesses and Economy – Subscribe to Business Today Magazine
OpenAI has expanded its Daybreak Program with a new specialised cybersecurity model, GPT-5.6-Cyber, and two new access tiers built for cybersecurity professionals. This will give cyber defenders access to more capable tools before attackers can weaponise similar AI systems at scale.
OpenAI Daybreak: Two Levels of Access
OpenAI’s Daybreak Program now offers two tiers of access: Daybreak Blue and Red. The Blue tier gives security researchers access to general-purpose advanced AI models, including GPT-5.6 Sol, to ensure safeguards such as finding software vulnerabilities, reviewing code, analysing malware, and responding to cybersecurity incidents.
Must read: OpenAI brings Premium Seats to ChatGPT Business with 5x more AI usage
The Daybreak Red offers greater specialised access, offering purpose-built cybersecurity models such as the new GPT-5.6-Cyber. This model can be used for vulnerability research, testing whether exploits actually work, and security testing.
GPT-5.6 Cyber: Fewer refusals, sharper exploit skills
The GPT-5.6-Cyber is built on GPT-5.6 Sol and is specifically trained for cybersecurity, such as finding zero-day vulnerabilities and developing exploit chains, but with fewer restrictions. During internal evaluations, OpenAI found that GPT-5.6-Cyber completed 95% of advanced cybersecurity requests. Whereas, safeguarded GPT-5.6 Sol completed only 1.5%. In addition, its predecessor, GPT-5.5-Cyber, completed around 57%.
Researchers are also using GPT-5.6 Cyber for vulnerability research. They also found two previously unknown vulnerabilities in V8, the JavaScript engine used by Chrome.
The vulnerabilities could reportedly be chained together to corrupt memory and escape Chrome's security sandbox. The vulnerabilities were reported to Google and subsequently fixed.
Must read: OpenAI rolls out GPT-5.6 Luna as default ChatGPT model, removes text chat limits for free tier users
It also helped find vulnerabilities across major mobile operating systems, a popular database, and an operating-system kernel, including hundreds of privilege-escalation vulnerabilities.
Unlike the standard GPT-5.6 Sol, which refuses sensitive cybersecurity requests due to fear of misuse, GPT-5.6-Cyber is designed to refuse fewer legitimate dual-use security requests. While OpenAI is reducing some model restrictions, it is also introducing added safeguards, such as hardware security keys, auto-review in Codex, and it also plans to expand monitoring of how the model is being used.
However, access to GPT-5.6 Cyber will remain limited to cybersecurity individuals and organisations with identity verification and signed legal attestations confirming authorised use.
For Unparalleled coverage of India's Businesses and Economy – Subscribe to Business Today Magazine