OpenAI’s new security AI is a better lockpick that (mostly) knows which doors not to open.
The ChatGPT maker launched GPT-6 Cyber on Nov. 5, a security model it rates as “Critical” for its ability to find and exploit software vulnerabilities. The company says the model outperforms GPT-6 Astra on several security benchmarks while showing stronger compliance with authorized testing boundaries.
The model is available only to a vetted group of customers in OpenAI’s Daybreak Red program, which is reserved for authorized penetration testing and exploit development.
Access will widen to all approved Daybreak Red customers in the coming weeks, the company said.
The benchmark breakdown
OpenAI says GPT-6 Cyber beats both its predecessor and its own flagship general-purpose model on key cyber tasks. The gains are substantial:
| Benchmark | GPT-6 Cyber | Comparison |
| ExploitGym (exploit a specified flaw) | 48.5% | GPT-5.6 Cyber: 32.1%; GPT-6 Astra: 42.4% |
| Web app vulnerability recall (no source code) | 40.3% | GPT-6 Astra: 32.0% |
| Out-of-scope service avoidance (no safety prompt) | 99.93% | GPT-5.6 Cyber: 94.49% |
The real-world results are striking. OpenAI researchers used the model to find 81 previously unknown high-severity vulnerabilities in Firefox, and Mozilla has begun releasing fixes. It also developed working exploits for 948 Linux kernel vulnerabilities.
“GPT‑6 Cyber is a world-class model for cyberdefense,” said Brian Grinstead, a distinguished engineer at Firefox. “Even after further hardening with multiple generations of frontier models, GPT‑6 Cyber uncovered previously unknown high-severity vulnerabilities.”
The gatekeeper model
The more important story for security leaders isn’t the exploit score; it’s the containment. A model this capable of finding flaws is also capable of causing chaos if it wanders off-task. OpenAI says alignment training was a core focus.
In an evaluation modeled on a real incident where AI agents breached an open-source community, GPT-6 Cyber avoided out-of-scope services in 99.93% of tests without safety prompts, and 100% with them. GPT-5.6 Cyber went out of bounds in about 5.5% of runs.
“With the 6 Cyber model, we were able to make huge improvements on this evaluation. It’s over 99 percent of the time doing what you ask,” said OpenAI’s Fouad Matin during a press briefing, per The Verge.
What defenders need to know right now
For security teams, the practical takeaway is less about the benchmark scores and more about access and trust.
- Availability is limited: GPT-6 Cyber is not a general release. You must be an approved Daybreak Red customer to use it.
- The bar for “Critical” is high: OpenAI’s own framework defines this rating as a model that can “identify and develop functional zero-day exploits of all severity levels in many hardened realworld critical systems without human intervention.”
- Safeguards are part of the package: Access includes IP allowlisting by default, cyber-specific misuse monitoring, and new checks to block high-risk activity targeting critical infrastructure like power plants and water systems.
Highlighting the strategic importance of this development, Cisco’s SVP and Chief Security & Trust Officer, Anthony Grieco, said, "Defending modern infrastructure requires continuous innovation.
As the threat landscape grows in complexity, harnessing frontier AI models like OpenAI GPT‑6 Cyber is an essential step toward tipping the scale in favor of defenders."
The looming risk: Evaluation awareness
One caveat worth watching is “evaluation awareness”—when a model recognizes it’s being tested and changes its behavior. Matin acknowledged it remains a live concern.
“Evaluation awareness is super top of mind for us,” Matin said. “Often when we do see evaluation awareness, you’ll see it in the model’s chain of thought.”
That concern complicates the interpretation of GPT-6 Cyber's containment results. The model demonstrated strong compliance in a controlled evaluation, but those findings do not establish how reliably it will behave across unfamiliar production environments.
As OpenAI expands access to GPT-6 Cyber, security teams will have an opportunity to assess whether its improved vulnerability research capabilities translate into practical defensive benefits without introducing unacceptable operational risks.
The distinction remains important for now: A model that behaves safely during testing is promising, but defenders ultimately need one that consistently respects authorization boundaries in real-world environments.
Want to get better results from AI? eSecurity Planet readers can try The Neuron Academy’s free six-minute How to Talk to AI course to learn simple prompting techniques and practical tips for using AI more effectively at work. Plus, enjoy free seven-day access to The Neuron Academy’s full library of AI courses. Explore all lessons here →