Sunday, November 8, 2026 --:--:-- ET
New York 62°F Partly cloudy Los Angeles 78°F Sunny Chicago 55°F Windy
Subscribe
U.S. Edition Est. 2026 Nov. 8, 2026

US News America

Independent reporting for a changing America

Breaking
USCIS Memo Adds Child Support and Public Benefits to Citizenship Character Test Burrow Throws for 271 Yards and Three Touchdowns as Bengals Beat Falcons 34-17 in Madrid Flash Flood Warning Issued in South Texas as Slow Storm Brings a Second Round of Heavy Rain Three Oil Tankers on Fire in Strait of Hormuz After Blasts, Iran’s Revolutionary Guards Say Trump Orders Justice Department to Look Into Democrats Over Threats of Impeachment
Technology

OpenAI Rates GPT-6 Cyber Its Highest Risk Level After the Model Built Working Exploits for 948 Linux Flaws

OpenAI released GPT-6 Cyber, a model built for authorized security research, and gave it a Critical cybersecurity rating after it built working exploits for 948 Linux kernel flaws and found 81 previously unknown high-severity Firefox bugs.

By admin Nov. 7, 2026 · 12:39 PM ET 2 min read

Two analysts seen from behind at desks in a dark security operations center, facing a wall of large monitors showing blue network diagrams and charts.

OpenAI has released GPT-6 Cyber, a model built for authorized security research, and rated it “Critical” for cybersecurity under the company’s Preparedness Framework, the highest rating in that category. The rating is one step above the “High” rating given to the previous model, GPT-5.6 Cyber.

A computer monitor showing lines of code and a circular security graphic in blue and teal.
A computer monitor showing lines of code and a circular security graphic in blue and teal.

The company said GPT-6 Cyber is available first to a small group of approved customers in Daybreak Red, its program for authorized offensive security work, through the API and Codex. Access will expand to more approved Daybreak Red customers over the coming weeks, according to OpenAI.

OpenAI said the model scored 48.5% on ExploitGym, a test of whether a model can turn a known flaw into a working exploit, against 32.1% for GPT-5.6 Cyber and 42.4% for GPT-6 Astra. In a web penetration test, it found 40.3% of the flaws, against 32% for Astra. In a separate test of remote attacks, it completed 95% of the tasks while Astra refused them all.

What the model found in Linux and Firefox

OpenAI said its researchers used GPT-6 Cyber to reproduce 1,724 Linux kernel vulnerabilities and build working exploits for 948 of them, including 588 flaws that were previously unknown. The company said it sent 304 critical findings to Linux maintainers along with proposed patches. OpenAI also said the Linux work cost about $1.2 million in API credits.

A corridor between rows of server racks with blinking green and blue lights in a data center.
A corridor between rows of server racks with blinking green and blue lights in a data center.

In Firefox, the model found 81 previously unknown flaws that Mozilla classed as high severity, and it showed that 30 known issues were worse than reported, according to OpenAI. Mozilla has begun releasing fixes for the flaws.

OpenAI also said it used GPT-6 Cyber with other models to develop and validate exploits for a flaw in Brotli, a compression library used by Node.js, that could allow remote code execution in downstream applications.

Limits on who can use it

The same skills that make the model useful to defenders could also help attackers, and OpenAI said it is limiting access for that reason. Daybreak Red API access requires IP allowlisting by default, and the company said it monitors for misuse. In company tests, the model stayed within the limits of an authorized task in 99.93% of cases without a deployment prompt, and in every test with one, against 94.49% for GPT-5.6 Cyber.

A glowing blue wireframe padlock hologram floating above a circuit board.
A glowing blue wireframe padlock hologram floating above a circuit board.

OpenAI also announced a $1 billion grant program, Daybreak for Frontline Defenders, to subsidize access for teams that protect power grids, water systems and other critical infrastructure. A separate group, Daybreak Red for Critical Infrastructure, lets approved defenders run authorized tests on critical systems.

All of the performance figures in this article come from OpenAI’s own announcement and system card, and they have not been checked by an outside lab. SecurityAffairs and Security Boulevard both note that the numbers should be read as the company’s claims.

Reporting based on coverage by SecurityAffairs and Security Boulevard.

Related Stories