OpenAI on Thursday launched GPT-6 Cyber, a specialised cybersecurity model it ranks as having ‘critical’ ability to identify vulnerabilities and develop exploits, up from ‘high’ for its predecessor. Built on GPT-6 Astra, the model is initially available only to a select group of approved Daybreak Red customers; Daybreak Red is for authorized red teaming, penetration testing and exploit validation.
OpenAI said its own researchers used the new model to uncover flaws and build exploits in Firefox and the Linux kernel, while security teams at Cisco and CodeRabbit already rely on it for code review and vulnerability hunting. The company plans to eventually open the model to all Daybreak Red users and expand the Daybreak program to more security researchers, organisations and open-source maintainers.
“It is clear that we have a limited window to use frontier models to secure software,” OpenAI wrote.
GPT-6 Cyber is a specialised version of GPT-6 Astra, trained with reinforcement learning to reproduce crashes, develop exploits and find vulnerabilities in code while rewarding it for staying within the user’s authorised scope and reducing unnecessary refusals.
On ExploitGym, a benchmark that tests whether models can exploit a specified vulnerability on a remote target, GPT-6 Cyber scored 48.5%, compared with GPT-5.6 Cyber’s 32.1%, OpenAI said. For finding vulnerabilities in web applications without source-code access, the new model scored 40.3%, compared with 32% for the widely available GPT-6 Astra. On ExploitBench, which tests exploit development from known vulnerabilities, OpenAI reported a 100% mean capability score, though it noted the benchmark includes vulnerabilities dating back to 2023 that may have appeared in training data, potentially inflating the results.
OpenAI also introduced RemoteExploitBench, a benchmark that asks models to exploit known vulnerabilities in applications such as GitLab, PostgreSQL and Jenkins in a controlled environment.
The company tested GPT-6 Cyber’s ability to stay within its assigned task using a new test built with the Hugging Face incident in mind: the ExploitGym Honeypot. It gives the model a difficult exploitation task inside a sandbox that also includes helper services OpenAI deliberately left easy to break into, hacking them would be a shortcut, but they are outside the task’s scope.
Without a deployment safety prompt, GPT-6 Cyber left those out-of-scope services alone in 99.93% of tests. With the prompt, it never touched them. Under the same conditions, GPT-5.6 Cyber went after those services in 5.51% of runs, and the general-purpose GPT-5.6 Sol did so in almost half of them. GPT-6 Astra, the base model, never took the bait.
“This improvement in alignment reflects dedicated training to stay within authorized scope, respect restrictions, and report results accurately,” OpenAI wrote.
Analysis
Why This Matters
- The model’s critical offensive capability could accelerate both security research and offensive cyber operations, depending on who uses it.
- OpenAI’s focus on alignment and staying within scope sets a precedent for deploying powerful hacking AIs without unintended side effects.
- Expansion to open-source maintainers and the broader Daybreak program may help secure widely used software at scale.
Background
OpenAI introduced its first Cyber model series to provide specialised AI tools for cybersecurity professionals, offering both offensive (Red) and defensive (Blue) variants. The new GPT-6 Cyber builds on the company's latest general-purpose model, GPT-6 Astra, and incorporates extensive reinforcement learning to balance exploit capability with task boundaries. The Daybreak program, through which the model is accessed, is OpenAI's controlled distribution channel for security teams and researchers.
Key Perspectives
OpenAI: Says frontier models offer a limited window to secure software, and that GPT-6 Cyber represents a major leap in security research with improved alignment that respects authorised scope and reduces unnecessary refusals.
What to Watch
- How quickly OpenAI expands Daybreak Red access beyond the initial select group to all approved users and open-source maintainers.
- Real-world adoption by major security teams, including whether Cisco and CodeRabbit extend their use beyond the pilot stage.
- Independent validation of benchmarks, particularly the 100% ExploitBench score and its potential inflation from training data.
Checked
12 claims in this story checked against 1 source: 10 supported.
Show each claim
- SupportedOpenAI ranks GPT-6 Cyber’s ability to identify vulnerabilities and develop exploits as “critical,” up from “high” for its predecessor.The New Stack: “The company ranks GPT-6 Cyber’s ability to identify vulnerabilities and develop exploits as ‘critical,’ up from ‘high’ for its predecessor.”
- SupportedOpenAI launched GPT-6 Cyber on Thursday.The New Stack: “OpenAI on Thursday launched GPT-6 Cyber”
- Partly supportedGPT-6 Cyber exploits flaws in Firefox and the Linux kernel.The New Stack: “It notes that its own researchers used the new model to uncover flaws and build exploits in Firefox and the Linux kernel”
- SupportedGPT-6 Cyber is initially available only to a select group of approved Daybreak Red customers.The New Stack: “GPT-6 Cyber will not be available to all developers but only to “a select group of approved Daybreak Red customers,” OpenAI says.”
- SupportedSecurity teams at Cisco and CodeRabbit already rely on GPT-6 Cyber for code review and vulnerability hunting.The New Stack: “security teams at Cisco and CodeRabbit already rely on it for code review and vulnerability hunting.”
- SupportedOpenAI wrote, “It is clear that we have a limited window to use frontier models to secure software.”The New Stack: ““It is clear that we have a limited window to use frontier models to secure software ,” OpenAI writes.”
- SupportedGPT-6 Cyber was trained with reinforcement learning to reproduce crashes, develop exploits and find vulnerabilities, while rewarding it for staying within authorised scope and reducing unnecessary refusals.The New Stack: “OpenAI says it used reinforcement learning to train the model to reproduce crashes, develop exploits, and find vulnerabilities in code, while also rewarding it for staying within the user’s authorized scope and reducing unnecessary refusals.”
- SupportedOn ExploitGym, GPT-6 Cyber scored 48.5%, compared with 32.1% for GPT-5.6 Cyber.The New Stack: “GPT-6 Cyber scores 48.5%, compared with GPT-5.6 Cyber’s 32.1%, OpenAI says.”
- SupportedFor finding web-application vulnerabilities without source-code access, GPT-6 Cyber scored 40.3%, compared with 32% for GPT-6 Astra.The New Stack: “For finding vulnerabilities in web applications without access to their source code, the new model scores 40.3%, compared with 32% for GPT-6 Astra”
- SupportedOpenAI reported a 100% mean capability score for GPT-6 Cyber on ExploitBench, while noting the benchmark’s older vulnerabilities may have appeared in training data and inflated the result.The New Stack: “On ExploitBench, which tests exploit development using known vulnerabilities, OpenAI reports a 100% mean capability score. The company notes, though, that the benchmark includes vulnerabilities dating back to 2023 and those examples may have appeared in training data, potentially inflating the results.”
- Partly supportedGPT-6 Cyber stayed within scope in 99.93% of tests.The New Stack: “Without a deployment safety prompt, GPT-6 Cyber left those services alone in 99.93% of tests.”
- SupportedOpenAI wrote, “This improvement in alignment reflects dedicated training to stay within authorized scope, respect restrictions, and report results accurately.”The New Stack: ““This improvement in alignment reflects dedicated training to stay within authorized scope, respect restrictions, and report results accurately,” OpenAI writes.”
