OpenAI has introduced GPT-6 Cyber, a specialised cybersecurity model designed to identify and validate exploitable vulnerabilities while following the boundaries and permissions established by security researchers.
The new model delivers significant improvements over GPT-5.6 Cyber across key cybersecurity benchmarks, including a 51% relative improvement on ExploitGym. It is already supporting real-world security research, helping uncover and validate thousands of exploits in critical software systems, including Linux and Firefox.
On ExploitGym, which evaluates a model’s ability to exploit a specified vulnerability in a remote target, GPT-6 Cyber achieved a score of 48.5%, compared with 32.1% for GPT-5.6 Cyber. It also demonstrated stronger alignment with assigned security boundaries. In tests designed to identify out-of-scope exploitation, the model avoided targeting unauthorised services in 99.93% of cases without deployment prompts and 100% with them.
OpenAI also introduced a new evaluation for offensive web penetration testing, where models identify vulnerabilities in web applications without access to their source code. GPT-6 Cyber achieved 40.3% vulnerability recall, compared with 32.0% for GPT-6 Astra.
Another benchmark, RemoteExploitBench, evaluates the ability to develop working exploits against known vulnerabilities in applications such as GitLab, PostgreSQL and Jenkins within controlled environments. In the initial evaluation, GPT-6 Astra with Daybreak Blue refused 100% of the tasks, while GPT-6 Cyber successfully completed 95%. Its only unsuccessful attempt involved an exploit that could not be executed remotely.
The model has also been assessed under OpenAI’s Preparedness Framework, where it is classified as Critical in cybersecurity, an increase from the High classification assigned to GPT-5.6 Cyber.
Real-world security research and vulnerability discovery
GPT-6 Cyber is already helping security researchers identify vulnerabilities, validate their impact and develop potential fixes.
- 81 previously unknown Firefox vulnerabilities: Researchers used the model to identify 81 new vulnerabilities classified as high severity by Mozilla. It also helped establish that 30 previously known issues were more severe than initially reported. OpenAI shared reproducible tests and proposed patches with Mozilla, which has started releasing fixes.
- 948 Linux kernel exploits: The model reproduced 1,724 vulnerabilities and developed working exploits for 948 of them. This is more than 5 times the 166 exploits developed by GPT-5.6 Sol using the same research framework. Following guidance from the Linux community, researchers disclosed 304 critical findings to maintainers with proposed patches.
- Critical remote code execution vulnerabilities: Researchers also used GPT-6 Cyber alongside other OpenAI models to develop and validate exploits for a critical vulnerability in Brotli, a compression library used by Node.js and other software. The findings demonstrated how the flaw could enable remote code execution in downstream applications.
Security teams at Cisco, Calif and CodeRabbit are also using the model for vulnerability research and code reviews. OpenAI has relied on it to identify and address weaknesses in its own infrastructure.
Daybreak expands access to advanced cybersecurity capabilities
OpenAI is working to expand Daybreak, its cybersecurity initiative, to organisations of different sizes, independent researchers and open-source maintainers.
The initiative includes different access configurations. Mainline models such as GPT-6 Astra support code reviews, vulnerability assessments and patch development. Daybreak Blue helps verified defenders validate vulnerabilities and fixes with fewer unnecessary refusals, while Daybreak Red enables more advanced exploit development and authorised security testing through GPT-6 Cyber.
OpenAI has also made alignment a central part of the model’s development. One evaluation, ExploitGym Honeypot, tests whether a model attempts to exploit supporting services outside its assigned task. GPT-6 Cyber avoided exploiting these services in 99.93% of cases without deployment safety prompts, compared with 94.49% for GPT-5.6 Cyber. With deployment prompts, it remained within the authorised scope in 100% of tests.
Stronger safeguards for responsible cybersecurity use
Alongside the release, OpenAI has introduced additional protections to reduce misuse while supporting legitimate security research. These include cyber-specific monitoring to check whether model actions match user permissions, targeted monitoring for emerging threats, Private Safety Processing and protection against stolen credentials.
Daybreak Red API access requires IP allowlisting by default, with limited exceptions. OpenAI also recommends sandboxing, action reviews and human oversight for high-risk activities.
Critical infrastructure protection is another major focus. OpenAI aims to prevent unauthorised activities involving power plants, water treatment facilities and operational technology and industrial control systems (OT/ICS).
Its $1 billion Daybreak for Frontline Defenders grant programme will subsidise access for resource-constrained teams, including government organisations and essential service providers. OpenAI is also introducing Daybreak Red for Critical Infrastructure, a new access group for approved defenders to conduct authorised testing of critical systems.
Additional safeguards have been developed to prevent activities that could compromise industrial controls, interfere with equipment, cause serious injuries or lead to major physical damage and life-threatening service disruptions.
To assess these protections, OpenAI partnered with Dragos, a global provider of extended operational technology cybersecurity. Dragos conducted controlled laboratory evaluations using cyber-permissive Daybreak Red models. The tests covered activities ranging from network discovery and asset identification to targeted reconnaissance, sensitive data extraction and potentially dangerous actions involving industrial devices.
OpenAI is also expanding partnerships with offensive security providers, purple teams, energy companies, critical infrastructure operators and national laboratories to evaluate the effectiveness of its models and safeguards.
The company plans to continue improving its security measures through collaboration with researchers and industry partners, while increasing investments in threat investigations involving OT/ICS environments.
Also read: Viksit Workforce for a Viksit Bharat
Do Follow: The Mainstream LinkedIn | The Mainstream Facebook | The Mainstream Youtube | The Mainstream Twitter
About us:
The Mainstream is a premier platform delivering the latest updates and informed perspectives across the technology business and cyber landscape. Built on research-driven, thought leadership and original intellectual property, The Mainstream also curates summits & conferences that convene decision makers to explore how technology reshapes industries and leadership. With a growing presence in India and globally across the Middle East, Africa, ASEAN, the USA, the UK and Australia, The Mainstream carries a vision to bring the latest happenings and insights to 8.2 billion people and to place technology at the centre of conversation for leaders navigating the future.