OpenAI Group PBC today released GPT-6 Cyber, a version of its GPT-6 Astra model designed specifically for authorized vulnerability research.
The new model addresses an issue that OpenAI’s own security work uncovered this summer. In July, agents working in the company’s test environments broke into Hugging Face Inc. while performing an assigned task. GPT-6 Cyber was trained to stay within the targets authorized by a researcher, and the security test that OpenAI created for the launch is based on this incident.
This test, ExploitGym Honeypot, gives the model a difficult task and surrounds it with supporting services that are deliberately easy to compromise. Attacking one of these services is considered a scope violation.
Because there were no safety instructions for use, GPT-6 Cyber went out of bounds in 0.07% of scored runs. With its predecessor, GPT-5.6-Cyber, this was the case in 5.51% of cases. Adding the instructions brought the observed rate of GPT-6 Cyber to zero.
OpenAI rates the new model “Critical,” the highest cybersecurity level in its Preparedness Framework and the same classification Astra received when it launched in September. GPT-5.6-Cyber is rated “High”. The company says GPT-6 outperforms Cyber Astra in most cyber tasks, and an early research campaign on Firefox uncovered 81 previously unknown vulnerabilities discovered by Mozilla Corp. were classified as “high severity”. Mozilla has since delivered fixes for many of these.
Irregular, an AI security lab mentioned in the system map, approached the gap with Astra and found that the two are largely comparable, aside from better GPT-6 Cyber results on long multi-stage tasks. None of the models progressed past fully hardened targets in OpenAI’s own challenge set.
GPT-6 Cyber is the fourth cyber-specific model that OpenAI has released this year. Like the first, GPT-5.4-Cyber, it follows an application process. Initial access is limited to a select group of approved customers in Daybreak Red, the tier of OpenAI’s Daybreak program intended for offensive work such as penetration testing and red teaming. These customers reach the model through the company’s application programming interface and codex, and API requests must come from approved networks by default.
When it launched Gemini 4 Argon in September, Google LLC also put vetted defenders first. A week later, Anthropic PBC converted its cyber verification program to three levels of access.
OpenAI said GPT-6 Cyber will reach all Daybreak Red customers in the coming weeks and GPT-5.6 Cyber is now open to every customer in the tier. Critical Infrastructure Defenders will have their own Daybreak Red group. On the defensive side, the company is expanding early Astra access to more organizations and individual security researchers via Daybreak Blue.
Image: OpenAI/GPT image 2.5
Support our mission to keep content open and free by interacting with theCUBE community. Join theCUBE Alumni Trust Networkwhere technology leaders connect, share information and create opportunities.
- Over 15 million viewers of theCUBE videosto spark conversations about AI, cloud, cybersecurity and more
- Over 11.4k theCUBE alumni — Connect with more than 11,400 technology and business leaders shaping the future through a unique, trusted network
Founded by tech visionaries John Furrier and Dave Vellante, SiliconANGLE Media has built a dynamic ecosystem of industry-leading digital media brands, reaching over 15 million elite tech professionals. Our new proprietary theCUBE AI Video Cloud is a game-changer in audience engagement, leveraging the theCUBEai.com neural network to help technology companies make data-driven decisions and stay at the forefront of the industry conversation.
