
Artificial intelligence has reached a new threshold in information security with OpenAI’s announcement of the GPT-6 Cyber model. The new model demonstrates advanced static and dynamic analysis capabilities, outperforming previous generations in vulnerability exploitation benchmarks. The tool has sparked intense debates within the development and systems engineering community due to its effectiveness in auditing complex and legacy codebases.
What changed and new features
According to the disclosed announcements, GPT-6 Cyber has been classified with Critical-level cybersecurity capabilities under OpenAI’s Preparedness Framework, advancing from the previous model’s High tier. In automated security audits, the model identified 81 previously unknown high-severity vulnerabilities in the Firefox browser and assisted in building patches, many of which have already been released by Mozilla. In Linux kernel scopes, the performance was even broader: the AI mapped 2,308 potential flaws, validated the impact of 1,724 of them, and generated functional local privilege escalation exploits for 948 distinct vulnerabilities, allowing unprivileged users to achieve root access. In ExploitGym benchmark comparative tests, the new model achieved success in 48.5% of scenarios, outperforming the 32.1% registered by the previous GPT-5.6 Cyber version.
Practical impact for developers
For software engineers, open-source maintainers, and defensive security teams, the arrival of models like GPT-6 Cyber demands an immediate shift in response speed and risk mitigation posture. Because the model is not publicly available — with initial access restricted to selected customers in OpenAI’s Daybreak Red program — temporal asymmetry becomes the primary challenge. Organizations face a limited window to utilize automated frontier tools to audit and harden their software before similar capabilities become widely accessible to malicious actors. Automation in flaw discovery and validation drastically reduces the time between the introduction of a code bug and the creation of an active exploit routine.
Limitations and what to evaluate
Despite expressive results in benchmarks and audits of large projects, adopting artificial intelligence in offensive and defensive testing brings severe operational and ethical constraints. Segmented access via programs like Daybreak Red, focused on authorized testing, and Daybreak Blue, aimed at defenders investigating vulnerabilities, highlights OpenAI’s attempt to control unregulated proliferation. However, recent ecosystem incidents show that autonomous agents and automated pentest frameworks are already employed by cybercriminals to bypass defenses at scale. Evaluating dependency integrity, hardening continuous integration environments, and adopting rigorous code reviews have become mandatory requirements to mitigate the risk of AI-accelerated exploitation.
What to expect next
The planned expansion of Daybreak programs to smaller organizations, independent researchers, and open-source maintainers signals an attempt to democratize access to advanced auditing tools. Nevertheless, the AI-driven cyber arms race will continue to pressure engineering teams. The priority in upcoming software development cycles must be automated remediation and the implementation of architectures resilient to flaws discovered at systemic speed.