Skip to content
Neural Daily

ModelsNews

GPT-6 Cyber: OpenAI's new cybersecurity model

🇭🇺 Olvasd el magyarul →

OpenAI's GPT-6 Cyber scores 48.5% on ExploitGym, found 81 new Firefox vulnerabilities and almost always stays within its authorized scope.

· 2 min read

OpenAI announcement graphic: GPT-6 Cyber text on a dark background with code symbols
Image: OpenAI · Press kit · Official announcement image from openai.com · Source

What happened?

On November 5, OpenAI introduced GPT-6 Cyber (OpenAI). The successor to GPT-5.6 Cyber finds and validates vulnerabilities, develops exploits and proposes fixes while staying within the scope a researcher sets. It is initially available through the API and Codex to a select group of Daybreak Red customers, with access expanding to all Daybreak Red customers soon.

How much better is it?

  • ExploitGym: 48.5%, up from 32.1% for GPT-5.6 Cyber. That is a 51% relative improvement.
  • Web pentesting without source code: 40.3% vulnerability recall, versus 32.0% for GPT-6 Astra.
  • RemoteExploitBench: completed 95% of tasks, while GPT-6 Astra under Daybreak Blue refused all of them.
  • Linux kernel: reproduced 1,724 vulnerabilities and built working exploits for 948. GPT-5.6 Sol managed 166 in the same harness.
  • Firefox: 81 novel vulnerabilities rated high severity by Mozilla, which has already shipped fixes for many of them.

OpenAI classifies the model as Critical in cybersecurity under its Preparedness Framework. GPT-5.6 Cyber was rated High.

Staying in scope

OpenAI used dedicated training to keep the model from attacking services outside its task. In the ExploitGym Honeypot evaluation, deliberately easy-to-exploit side services were placed in the environment. Without safety prompts, GPT-6 Cyber left them alone in 99.93% of cases (GPT-5.6 Cyber: 94.49%), and in 100% with its deployment prompt.

Stronger safeguards ship alongside it: misuse monitoring, IP allowlisting by default for Daybreak Red API access and new blocks on requests targeting critical infrastructure such as industrial control systems. A new Daybreak Red for Critical Infrastructure group gives approved defenders of those systems room for authorized testing.

Who is using it?

According to OpenAI, Cisco, Calif and CodeRabbit already use it for vulnerability research and code review. On CodeRabbit's internal benchmark it found 66% more high and medium severity issues than the company's production model.

Why it matters

Frontier models now not only find bugs but prove them with working exploits, which shortens the patch cycle. Access stays narrow, though: you need to apply for Daybreak. Meanwhile GPT-5.6 Cyber is becoming available to all Daybreak Red customers.

    By Mikola Máté Patrik

    Produced with the help of AI, reviewed by an editor. How we make our articles

    Reviewed and edited by the Neural Daily desk. About

    Related articles