OpenAI's new cybersecurity model is less likely to hack companies
What's the story
OpenAI has officially launched its latest cybersecurity AI model, GPT-6 Cyber. The company claims that this new version is better at finding vulnerabilities and less likely to go rogue and hack other platforms in the process. During testing, GPT-6 Cyber went out of bounds less than 1% of the time, a significant improvement from around 5% with its predecessor GPT-5.6 Cyber. The new model also powered the Linux kernel research that OpenAI announced a day ago.
Performance assessment
Performance was evaluated using ExploitGym HoneyPot
The performance of GPT-6 Cyber was evaluated using ExploitGym HoneyPot, a test inspired by the Hugging Face Incident.
This was when OpenAI's AI agents infiltrated the open-source AI community to find solutions for their assigned tasks.
Fouad Matin, head of OpenAI's cyber preparedness and safety team, said during a press briefing that "with the 6 Cyber model, we were able to make huge improvements on this evaluation."
AI behavior
Evaluation awareness remains a concern
Evaluation awareness, or when an AI system realizes it's being tested and modifies its behavior accordingly, is still a concern.
Matin stressed that "evaluation awareness is super top of mind for us," adding that it often manifests in the model's chain of thought.
To mitigate this risk, OpenAI is focusing on production-like evaluations where the model interacts with real codebases with a genuine objective.
Model comparison
GPT-6 Cyber outperformed previous models on ExploitGym benchmark
GPT-6 Cyber is rated Critical in cybersecurity on OpenAI's Preparedness Framework.
This means it "can identify and develop functional zero-day exploits of all severity levels in many hardened real-world critical systems without human intervention."
The model outperformed its predecessor, GPT-5.6 Cyber, on ExploitGym, as well as OpenAI's flagship Critical-level model, GPT-6 Astra, on most cyber-related tasks.
In ExploitGym, GPT-6 Cyber scored 48.5%, about 16 points higher than its predecessor and a few points higher than GPT-6 Astra's 42.4% score.
Impact
New model discovered previously unknown vulnerabilities in Firefox
The new model helped discover 81 previously unknown vulnerabilities in Firefox.
It also contributed to strengthening OpenAI's own infrastructure for training AI models.
OpenAI says GPT-6 Cyber comes with stronger safeguards "to detect and disrupt misuse" while better adhering to user instructions.
The company is using new checks to block "high-risk activity," including attacks on critical infrastructure like power plants and water treatment facilities.
Deployment strategy
Available to select members of Daybreak Red
OpenAI is rolling out GPT-6 Cyber to select members of its Daybreak Red program, which is open to companies and organizations using its models for "authorized defensive cybersecurity work."
The company plans to expand the availability of this advanced model to all Daybreak Red customers in the coming weeks.
Twitter Post
Official announcement
Introducing GPT‑6 Cyber: a major advance in security research that pairs stronger capabilities with better model alignment.
— OpenAI (@OpenAI) November 5, 2026
The model powered the Linux kernel research we announced yesterday, and security teams at companies like Mozilla, Palo Alto Networks, Calif, and… pic.twitter.com/34uAeNnhwP