OpenAI has unveiled GPT-6 Cyber, a groundbreaking AI model focused on enhancing cybersecurity. The model has impressively identified 81 new high-severity vulnerabilities in Firefox and successfully created exploits for 948 bugs in the Linux kernel. By improving its capability to adhere to defined security protocols, GPT-6 Cyber surpasses its predecessors in both scope and precision.

Released on November 4, 2026, GPT-6 Cyber is currently accessible to a select group of Daybreak Red customers through OpenAI's API and Codex platform. Its performance on various security benchmarks, such as ExploitGym and RemoteExploitBench, demonstrates remarkable advancements over previous models. GPT-6 Cyber now holds a Critical rating under OpenAI's Preparedness Framework, a step up from the High rating of GPT-5.6 Cyber.

A significant concern with AI in cybersecurity is the potential for models to stray from their intended tasks. However, tests with ExploitGym Honeypot show that GPT-6 Cyber remains focused, staying within scope 99.93% of the time without safety prompts, and 100% with them. This focus is achieved through meticulous programming and reasoning prompts that caution against exploring unauthorized credentials.

OpenAI has provided Mozilla with proof and test cases for the newly discovered Firefox flaws, many of which involve use-after-free memory errors. These findings are detailed in Mozilla's advisory MFSA 2026-118, which accompanies the release of Firefox version 159. Similarly, the model's contributions to Linux security are documented in OpenAI's report, 'A thousand fewer paths to root,' which outlines the extensive auditing process and the resulting vulnerability findings.

OpenAI has taken significant steps to ensure the responsible use of GPT-6 Cyber. The company has implemented monitoring to ensure model actions align with user permissions, and it has introduced a Daybreak Red for Critical Infrastructure group to enhance security for vital services. These efforts are complemented by OpenAI's emphasis on prompt patching, rescanning old code, and utilizing sandboxing techniques to mitigate risks.