OpenAI released GPT-6 Cyber on Thursday to a small group of vetted customers, rating the model Critical in cybersecurity under its safety framework. The company says the model finds vulnerabilities faster and stays within authorized boundaries more reliably than its predecessor. Security teams at Cisco, Calif, and CodeRabbit are already using it for research and code review, according to OpenAI.
OpenAI Launches GPT-6 Cyber With Restricted Access and Scope Controls
Reuters reported that OpenAI released GPT-6 Cyber on Thursday to a small group of vetted customers. The company rated the model Critical in cybersecurity under its Preparedness Framework safety system. OpenAI says the model is meant to help researchers find flaws and fix them before attackers act.
a statement describes GPT-6 Cyber as a cyber-specialized variant of GPT-6 Astra for authorized security work. The variant covers vulnerability discovery, exploit development, and web penetration testing, the system card says. OpenAI says specialized post-training aims to improve capabilities and reduce refusals for legitimate tasks.
OpenAI Releases GPT-6 Cyber, and Promises This One Stays Within Rules of Engagement
a statement says GPT-6 Cyber follows task boundaries more reliably than GPT-5.6 Cyber. In alignment testing, the model avoided exploiting out-of-scope services in 99.93% of tests without its deployment prompt. With that prompt, the model avoided out-of-scope exploitation in 100% of tests, OpenAI says.
The Verge reported that the model went out of bounds less than 1% of the time during testing. The Verge said that compared with around 5% for GPT-5.6 Cyber in earlier testing. OpenAI attributes the change to dedicated training to stay within authorized scope, the Verge reported.
OpenAI Debuts GPT-6 Cyber, Touting Better Performance and Safety Guardrails
OpenAI says GPT-6 Cyber outperforms GPT-5.6 Cyber on key cybersecurity benchmarks, including ExploitGym. On ExploitGym, which tests exploitation of a specified vulnerability, GPT-6 Cyber scores 48.5%, OpenAI says. That compares with 32.1% for GPT-5.6 Cyber on the same benchmark, according to the company.
OpenAI describes the benchmark result as a 51% relative improvement over GPT-5.6 Cyber on ExploitGym. The company also says the model shows improved alignment compared with GPT-5.6 Cyber. GPT-6 Cyber outperforms GPT-6 Astra on most cyber tasks, even when Astra does not refuse, OpenAI says.
Linux and Firefox Findings Test the Model on Real Code
OpenAI says researchers used GPT-6 Cyber to develop 948 privilege escalation exploits in Linux. The company says the exploits came from security-critical codebases that had extensive prior testing with frontier models. Human researchers reviewed findings before reporting vulnerabilities and proposed fixes, according to the brief and OpenAI.
The model also identified 81 previously unknown high-severity vulnerabilities in Firefox, OpenAI says. Mozilla has released fixes for many of those Firefox vulnerabilities, according to OpenAI. The model helped develop candidate patches for review by maintainers, OpenAI says.
Restricted Access Through Daybreak Red and Verified Defenders
Access to GPT-6 Cyber is restricted through Daybreak Red, OpenAI says in its system card. Daybreak Red requires identity verification, use restrictions, and organizational security measures. Those measures include single sign-on, multi-factor authentication, and API IP allowlisting, the system card says.
OpenAI says it aims to expand Daybreak for organizations, independent researchers, and open-source maintainers. Daybreak Blue helps verified defenders validate vulnerabilities and fixes with fewer unnecessary refusals, OpenAI says. Daybreak Red enables deeper exploit development and authorized security testing with GPT-6 Cyber, the company says.
Questions and answers
What is GPT-6 Cyber? GPT-6 Cyber is a cyber-specialized variant of GPT-6 Astra, OpenAI says. It is designed for authorized security work such as vulnerability discovery and exploit development. It is classified as Critical in cybersecurity under OpenAI’s Preparedness Framework.
Who can use GPT-6 Cyber? Access is restricted through Daybreak Red for verified users, OpenAI says. The program requires identity verification, use restrictions, and organizational security requirements. API customers using Zero Data Retention generally require Private Safety Processing, the system card says.
How does GPT-6 Cyber perform on ExploitGym? GPT-6 Cyber scores 48.5% on ExploitGym, OpenAI says. GPT-5.6 Cyber scored 32.1% on the same benchmark, according to OpenAI. OpenAI describes the change as a 51% relative improvement for GPT-6 Cyber.
How often did the model stay within scope? GPT-6 Cyber avoided exploiting out-of-scope services in 99.93% of tests without its deployment prompt, OpenAI says. With the deployment prompt, the model avoided out-of-scope exploitation in 100% of tests. The Verge reported the out-of-bounds rate was less than 1% in testing.
What real systems has the model tested? OpenAI says researchers developed 948 privilege escalation exploits in Linux using the model. The model also identified 81 previously unknown high-severity vulnerabilities in Firefox, OpenAI says. Mozilla has released fixes for many of those Firefox vulnerabilities, according to OpenAI.
What difference does Daybreak make? Daybreak Blue helps verified defenders validate vulnerabilities and fixes with fewer refusals, OpenAI says. Daybreak Red enables deeper exploit development and authorized security testing with GPT-6 Cyber. Mainline models such as GPT-6 Astra already support code review, triage, and patch development, OpenAI says.
OpenAI says the model is already helping secure real-world software in Linux and Firefox. The company says security teams face a limited window to use frontier models to secure software.




