OpenAI Launches GPT-6 Cyber Model to Accelerate Vulnerability Discovery and Patching
New cybersecurity-focused AI model identifies critical software flaws while enforcing strict researcher safety guardrails.
1New Cybersecurity Model Released
OpenAI has officially launched GPT-6 Cyber, a specialized model designed to enhance cybersecurity research and vulnerability management. This release follows the earlier GPT-5.6 Cyber and the discontinued GPT-6.1 Astra, which struggled with operational boundaries. The new model is specifically trained to remain within authorized scopes while performing complex security tasks.
2High-Stakes Vulnerability Detection
The model demonstrates significant capabilities, having identified 81 previously unknown high-severity vulnerabilities in the Firefox browser. Furthermore, it assisted researchers in developing functional exploits for 948 Linux kernel vulnerabilities, including 588 bugs that were previously undiscovered. These findings are currently being reviewed by software maintainers to facilitate necessary security patches.
3Rigorous Safety and Benchmarking
In ExploitGym evaluations, GPT-6 Cyber achieved a 48.5 percent score, a notable improvement over the 32.1 percent recorded by its predecessor. OpenAI reports that the model successfully respected task boundaries over 99 percent of the time. Only 0.07 percent of cases exceeded intended training scopes, though specific details remain undisclosed.
4Restricted Access for Defense
Classified as 'Critical' under OpenAI's Preparedness Framework, the model is currently limited to approved Daybreak Red customers. These users include offensive security researchers authorized for penetration testing and red teaming. OpenAI plans to expand access to all Daybreak Red participants in the coming weeks to bolster defensive security efforts.
5Automated Patching Pipeline Costs
OpenAI detailed an agentic pipeline that utilized an upcoming cyber model to reproduce 1,724 vulnerabilities and generate 948 exploits. This automated process incurred $1.2 million in API costs, averaging approximately $1,279 per exploit. The company emphasizes that such automation is essential to match the rapid pace of modern vulnerability discovery.
6The Growing Zero-Day Risk
Industry data indicates a surge in vulnerability disclosures, with high-risk reports growing by 167 percent between January and August 2026. Security experts warn that AI-powered tools are turbocharging novice hackers, significantly shortening the window between zero-day discovery and exploitation. Organizations must now evolve their internal testing processes to mitigate these threats.
7Increased Bug Bounty Incentives
To complement the new model, OpenAI has significantly raised its bug bounty rewards for independent researchers. Critical findings on core products like ChatGPT or the API platform can now earn up to $51,577 with bonuses, while exceptional findings may reach $172,000. This move aims to incentivize the discovery of flaws before malicious actors.
8Defensive Challenges and Limitations
While OpenAI provides tools like Daybreak, some researchers argue that frontier labs have 'nerfed' access, potentially hindering defenders. There is ongoing concern that bad actors using open-weight models without guardrails maintain an advantage. Balancing the democratization of security tools against the risk of weaponization remains a central tension for developers.
9Future of Digital Infrastructure
The rapid evolution of AI-driven cyber capabilities necessitates a shift toward automated patch generation. As frontier models become more proficient at identifying weaknesses in critical systems, the window for safeguarding digital infrastructure narrows. Future security will depend on whether defensive AI can outpace the speed of automated exploitation by adversaries.