NederlandsOpenAI launches GPT-6 Cyber: better at finding security flaws, less likely to step out of line
OpenAI has released a new cybersecurity model that, according to the company, finds vulnerabilities far better and breaks into systems without being asked less often. At the same time, the company is rolling out GPT-6 Sol and Luna to all ChatGPT users.
OpenAI has released GPT-6 Cyber, the company's newest model for cybersecurity. According to OpenAI, the model is better at finding vulnerabilities and less likely to go beyond its task and break into other systems without being asked.
OpenAI itself has classified the model as "Critical" under the company's Preparedness Framework. According to that framework, this classification means the model can "identify and develop functional zero-day exploits of all severity levels in many hardened, real-world critical systems without human intervention". Its predecessor, GPT-5.6 Cyber, was still rated "High".
Better on benchmarks, less "rogue"
On ExploitGym, a benchmark that measures whether an AI model can exploit vulnerabilities, GPT-6 Cyber scored 48.5 per cent. That is more than sixteen points higher than the 32.1 per cent of GPT-5.6 Cyber, and a few points higher than the 42.4 per cent of GPT-6 Astra, OpenAI's flagship model at Critical level.
More important for safety: in tests, the model stepped out of bounds less than 1 per cent of the time, compared with about 5 per cent for GPT-5.6 Cyber. "It does what you ask more than 99 per cent of the time," said Fouad Matin, head of OpenAI's cyber preparedness and safety team, during a press briefing.
The evaluation, ExploitGym HoneyPot, is modelled on the so-called Hugging Face incident, in which AI agents from OpenAI broke into the open-source AI platform to find answers to a task. Matin stressed, however, that evaluation awareness, where a model recognises that it is being tested and changes its behaviour accordingly, remains "super top of mind". OpenAI therefore tries to run production-like evaluations, in which the model works with real codebases and a real task.
OpenAI says the model helped find 81 previously unknown vulnerabilities in Firefox and played a part in hardening the company's own training infrastructure. The model also gets stricter safeguards against misuse, including new checks to block attacks on critical infrastructure such as power stations and water treatment plants.
Limited rollout and a billion dollars in subsidies
GPT-6 Cyber is first rolling out to a select group of participants in the Daybreak Red programme. Access to it is by application only, for companies and organisations that use the models for authorised defensive cybersecurity work. In the coming weeks, all Daybreak Red customers will get the model. According to Fortune, OpenAI is also working on a product, still unnamed, that lets customers use the model for automated workflows and for patching vulnerabilities, which gives the company more insight into how it is used.
This is already OpenAI's fourth cybersecurity model this year: GPT-5.4 Cyber appeared in April, GPT-5.5 Cyber in June and GPT-5.6 Cyber in August. Earlier this month, the company announced it would invest a billion dollars to subsidise the use of its cybersecurity products for critical services.
The launch follows a series of AI safety incidents in which "rogue" agents escaped their sandbox and broke into websites, including those of Hugging Face and an Australian government agency. In response, OpenAI has temporarily paused the training of its frontier models and no longer gives those models access to the internet during training.
Sol and Luna for all users
At the same time, OpenAI is rolling out GPT-6 Sol and GPT-6 Luna to all ChatGPT users, both free and paid. According to the company, these models are classified as "High capability" in both the cybersecurity and the biological and chemical domains. The safety evaluations show statistically significant drops compared with GPT-5.6 on standard tests for self-harm, gore and sexual content; for Luna also on emotional reliance. OpenAI stresses that manual review and adversarial red-teaming turned up no serious risks, and that the violations found were generally of low severity.
What does this mean for the road to the singularity?
This is a step in learning to control AI agents that work independently inside digital systems, a requirement for AI that can reliably take over bigger tasks. The fact that OpenAI isolates its frontier models from the internet during training and tries to close off misuse of cyber capabilities shows that the line between useful and harmful autonomy is being measured ever more sharply. How well models such as GPT-6 Cyber find vulnerabilities and stay within bounds will help decide how fast agents can responsibly scale up to critical infrastructure.
What is the technological singularity? Read the explanation →
