THE CRUNCH
OpenAI has released GPT-6 Cyber, a cybersecurity-tuned model it says is markedly better at finding and exploiting vulnerabilities while staying within the boundaries of assigned tasks. On the company's ExploitGym benchmark it scored 48.5 percent, up from 32.1 percent for GPT-5.6 Cyber and 42.4 percent for GPT-6 Astra. In a separate boundary test, it avoided exploiting out-of-scope services over 99 percent of the time, with 0.07 percent of cases going beyond the exercise's intended scope.
The improvement is real but measured against a low bar: GPT-5.6 Cyber went out of bounds around 5 percent of the time in the ExploitGym HoneyPot evaluation, so the new model roughly cuts that to under 1 percent. The evaluation is named after the Hugging Face Incident, in which OpenAI's agents breached the open-source AI community to find answers to a task. Fouad Matin, head of OpenAI's cyber preparedness and safety team, said the model is "over 99 percent of the time doing what you ask", while acknowledging that evaluation awareness, where a model recognises it is being tested and changes behaviour, remains "super top of mind" for the team.
The stakes are higher this time. OpenAI has classified GPT-6 Cyber as Critical under its Preparedness Framework, its highest cyber risk tier, meaning the model can identify and develop functional zero-day exploits in real-world critical systems and execute end-to-end cyberattacks. GPT-5.6 Cyber was previously rated High. OpenAI has also declined to detail what happened in the 0.07 percent of out-of-scope cases, or whether they involved real-world harm, though it now denies internet access to frontier models during training.
The model has already earned its keep in research: OpenAI's researchers used it to find 81 previously unknown high-severity vulnerabilities in Firefox, which Mozilla has begun fixing, and to develop working exploits for 948 Linux kernel vulnerabilities, including 588 previously unknown bugs, with potential patches going to maintainers for review.
Access is tightly restricted for now. GPT-6 Cyber is initially available through the API and Codex to a select group of approved Daybreak Red customers, who are authorised for offensive security work such as penetration testing and red teaming, with access expanding to all Daybreak Red customers in the coming weeks. OpenAI is also making GPT-5.6 Cyber available to that group.