OpenAI has unveiled its latest AI model, GPT-6 Cyber, which is making significant strides in cybersecurity. This advanced model has identified 81 high-severity vulnerabilities in Firefox and developed working exploits for 948 issues in the Linux kernel. According to OpenAI, GPT-6 Cyber adheres more effectively to set boundaries compared to its predecessor models.
Launch and Performance Metrics
GPT-6 Cyber was launched on November 4, 2026, and is initially available to a select group of approved Daybreak Red clients via API and Codex. The model showed impressive performance in ExploitGym, where it demonstrated a 48.5% success rate in exploiting known flaws, significantly surpassing the 32.1% achieved by GPT-5.6 Cyber. Additionally, in a new web testing task, GPT-6 Cyber identified 40.3% of flaws, outperforming GPT-6 Astra’s 32.0%.
In the RemoteExploitBench, GPT-6 Cyber completed 95% of tasks, while Astra with Daybreak Blue opted out of all tasks. OpenAI has rated GPT-6 Cyber as Critical for cybersecurity within its Preparedness Framework, an upgrade from GPT-5.6 Cyber’s High rating.
Challenges and Safety Measures
A persistent challenge with security AI models is ensuring they stay within their intended scope. OpenAI addressed this with ExploitGym Honeypot, which simulates potential distractions. GPT-6 Cyber remained within scope 99.93% of the time without safety prompts and achieved 100% with them. The model’s reasoning notes frequently emphasize the importance of not straying from designated credentials.
OpenAI has provided Mozilla with evidence and test cases for 81 newly discovered high-severity bugs. Additionally, the model revealed that 30 known issues were more severe than previously reported. Mozilla’s advisory MFSA 2026-118 documents these findings in Firefox 159, released in late October.
Linux Kernel Exploits and Future Outlook
OpenAI’s comprehensive report titled “A thousand fewer paths to root” details the discoveries within the Linux kernel. The model evaluated 5,034 audit targets, generating 2,308 candidate findings and reproducing 1,724. Ultimately, GPT-6 Cyber built 948 working local privilege escalation exploits. This achievement is a substantial increase from the 166 exploits identified by GPT-5.6 Sol in July.
The financial implications of this run were significant, costing $1,213,430 in API credits. Market valuations of these exploits range between $18.9 million and $66.3 million. OpenAI has disclosed 304 critical findings, with 61 issues resolved and 43 patches staged within ten days.
OpenAI emphasizes swift patching, rescanning old code upon model improvements, and treating low-privilege accounts on unpatched systems as potential root access points. The company recommends sandboxing techniques and constant vigilance for cybersecurity teams.
Conclusion and Recommendations
OpenAI advises organizations to update their systems promptly and conduct regular security assessments. With the introduction of the Daybreak Red for Critical Infrastructure group and the $1 billion Daybreak for Frontline Defenders grant, OpenAI is stepping up its efforts to protect essential services like power plants and water systems. As AI technology continues to evolve, so must our approach to cybersecurity, ensuring infrastructures remain robust against emerging threats.
