OpenAI has unveiled GPT-6 Cyber, a cutting-edge security model that boasts significant improvements over its predecessor. This new model is designed to hunt down flaws without crossing the line into malicious behavior, and it’s already showing promising results in identifying vulnerabilities.
GPT-6 Cyber was tested on ExploitGym, which simulates real-world scenarios where an AI might be used to exploit a flaw in a remote target. The results are impressive: GPT-6 Cyber achieved 48.5% success rate, a marked improvement over GPT-5.6 Cyber’s 32.1%. This is not just a matter of brute force; OpenAI has refined the model’s training data to emphasize alignment with user intentions and respect for boundaries.
One of the key features of GPT-6 Cyber is its ability to detect vulnerabilities in complex systems like the Linux kernel. In testing, it reproduced 1,724 known vulnerabilities and even developed functional exploits for 948 of them. This level of performance is unprecedented, and OpenAI researchers are confident that this model will be a valuable tool in the fight against cyber threats.
But GPT-6 Cyber’s capabilities don’t stop there. When tasked with identifying flaws in Firefox, it successfully documented 81 new high-severity vulnerabilities, providing evidence and reproducible tests to back up its findings. Mozilla has already begun rolling out patches to address these issues.
OpenAI is not only releasing this powerful tool but also implementing safeguards to prevent its use for malicious purposes. The company is introducing monitoring mechanisms that check whether the model’s actions align with user permissions, as well as restrictions aimed at preventing GPT-6 Cyber from being used to compromise safety-critical systems like industrial controls.
The feedback loop between OpenAI and its partners has been instrumental in refining these safeguards. For instance, Dragos, an industrial cybersecurity firm, provided critical input during testing of the Daybreak Red models – a part of OpenAI’s Daybreak program focused on advanced exploit work and authorized testing. The result is a model that not only excels at identifying vulnerabilities but also operates within clearly defined boundaries.
“This skill floor is dropping for the reconnaissance and vulnerability research needed before an attack,” said Dragos chief AI officer Rob T. Lee, highlighting the significant impact GPT-6 Cyber could have on cybersecurity efforts.
While access to this powerful tool is currently limited to select approved customers of Daybreak Red, OpenAI plans to gradually open it up to all users in the future. A dedicated group for critical infrastructure defenders will also be established, allowing sanctioned testing of critical systems.
As we move forward with AI-powered security tools like GPT-6 Cyber, one key takeaway is clear: responsible development and deployment are crucial to ensuring these models don’t become liabilities rather than assets. By focusing on alignment, safety prompts, and safeguards, OpenAI has taken a significant step towards making AI-powered cybersecurity a reality – one that can be trusted to protect us from the threats we face online.
Source: SecurityWeek — 2026-11-06