Go real-time, for free

The news here runs 15 minutes behind. Sign up free and see every story the moment it is published.

Already have an account? Sign in

Raydn

delayed by 15 minutes - go real-time About API Sign in
EN
Raydn Sign in Back to the front page
AI and Tech

OpenAI’s launches GPT-6 Cyber, its strongest hacking model yet

The model finds and exploits software flaws far better than its predecessor, OpenAI says, while straying off target much less often. Only vetted Daybreak Red customers have access.

OpenAI released GPT-6 Cyber on Thursday, a cybersecurity model built on its GPT-6 Astra system and aimed at finding and validating software vulnerabilities. Access is limited to a select group of approved customers in Daybreak Red, the application-only tier of OpenAI's cyber defense program.

OpenAI says the model scored 48.5% on ExploitGym, a test of whether an AI can exploit a known flaw in a remote target. That compares with 32.1% for its August model, GPT-5.6 Cyber.

The company also says the model produced reproducible evidence for 81 previously unknown high-severity vulnerabilities in Firefox and helped build 948 exploits for the Linux kernel. OpenAI rates GPT-6 Cyber as Critical for cybersecurity under its Preparedness Framework, up from High for its predecessor. Those figures come from OpenAI and have not been independently verified.

In a honeypot test run without deployment safety instructions, OpenAI says GPT-6 Cyber went after an out-of-scope target in 0.07% of graded runs, against 5.51% for GPT-5.6 Cyber. In plain terms, the company is selling a much stronger offensive tool that stays inside the rules set by the researcher using it.

The launch also raises the stakes in the race with Anthropic, whose Mythos model and Project Glasswing set the pace in AI-driven security work. Pricing was not disclosed, and the model is not available in ChatGPT. OpenAI has a "Cyber at the Frontier" livestream set for November 12, where more details are expected.

© Raydn · raydn.us