ToolOpenAI

OpenAI launches GPT-6 Cyber to find and fix security flaws

Original title: GPT‑6 Cyber: Advancing the frontier of aligned cybersecurity agents

Plain language

What it is: OpenAI introduced GPT-6 Cyber, a new artificial intelligence model designed for cybersecurity. It helps find security weaknesses in software and create fixes for them.
How it works: The model is trained to find and test security vulnerabilities while strictly staying within the boundaries set by researchers. It can also suggest code patches to fix the discovered issues.
Where it is available: The model is available through OpenAI's Daybreak program, specifically in the Daybreak Red tier for authorized security testing. It is used by security teams at companies like Cisco, CodeRabbit, and Mozilla.
Example: Researchers used the model to find 81 previously unknown high-severity security vulnerabilities in the Firefox web browser and suggest patches for them.

What it is: OpenAI released GPT-6 Cyber, a cybersecurity-focused model trained for vulnerability discovery, exploit generation, and patch development. It is classified as Critical under OpenAI's Preparedness Framework.
How it works: It uses specialized training to improve alignment, allowing it to stay within authorized task boundaries and avoid out-of-scope targets. It outperforms previous models on benchmarks like ExploitGym and RemoteExploitBench.
Where it is available: It is deployed within OpenAI's Daybreak program, specifically under the Daybreak Red tier for authorized exploit development. Security teams at organizations like Cisco, Calif, and CodeRabbit are currently using it.
Example: The model successfully developed 948 working privilege escalation exploits for the Linux kernel and reproduced 1,724 vulnerabilities.

The headline and explanation on this page are based on OpenAI's post. The full details are in the original.

GPT‑6 Cyber: Advancing the frontier of aligned cybersecurity agents