OpenAI Releases GPT-6 Cyber to Vetted Security Teams After Model Finds 588 Unknown Linux Flaws, Rated Critical

SAN FRANCISCO — OpenAI has released a specialized artificial intelligence model designed to find and exploit software vulnerabilities, but it is limiting access to vetted security customers after rating the model's cyber capabilities at its highest risk level.

OpenAI released GPT-6 Cyber, a model built to find and exploit security flaws, to a small group of vetted customers on Thursday, rating it "Critical" in cybersecurity under its safety framework. That is up from "High" for its predecessor, GPT-5.6 Cyber.

GPT-6 Cyber is a version of OpenAI's GPT-6 Astra model trained for authorized vulnerability research.

Purpose

OpenAI said the model is meant to help security researchers find vulnerabilities, test how they could be exploited and fix flaws before attackers do, amid increasing attempts by malicious cyber actors to misuse AI.

The company acknowledged the risks. OpenAI said that as models become more capable and uncover more vulnerabilities, "we are also seeing more attempts to use them to cause harm."

Results

OpenAI says the model has already found serious flaws in widely used software. Its researchers used the model to find 81 previously unknown Firefox vulnerabilities that Mozilla classified as high severity, and it helped develop working exploits for 948 Linux kernel vulnerabilities, including 588 that were previously unknown. OpenAI disclosed 304 critical findings to maintainers along with proposed patches.

On OpenAI's ExploitGym test, which measures the ability to exploit specified vulnerabilities, GPT-6 Cyber scored 48.5%, compared with 32.1% for GPT-5.6 Cyber.

Independent verification is limited so far. One security outlet noted the benchmark results and bug fixes are vendor-reported and remain unverified.

Who can use it

GPT-6 Cyber is available to select customers of Daybreak Red, OpenAI's access tier for authorized security work such as penetration testing and exploit development, and access will expand to all approved Daybreak Red customers over the coming weeks.

Early users include major companies. The tool is already in use by security teams at Mozilla, Palo Alto Networks and Cisco.

Safeguards

OpenAI has added monitoring specific to cyber tasks, checks on whether model behavior matches a user's authorization, and limits on requests that could cause severe harm to critical infrastructure and essential services. API requests must come from approved networks by default.

The model appears better at staying within limits. Without deployment safety instructions, GPT-6 Cyber went out of bounds in 0.07% of graded runs, compared with 5.51% for GPT-5.6 Cyber.

OpenAI is also backing a $1 billion program to subsidize access for government agencies and essential-service providers with limited resources.

Competition

Google and Anthropic are also competing for security customers, with Anthropic's Glasswing program giving vetted users access to capabilities its public models might block.