OpenAI releases GPT-6 Cyber with Critical safety rating

Overview

Updated 51 minutes ago

OpenAI's new security model found 81 previously unknown high-severity flaws in Firefox. It also built working exploits for 948 Linux kernel vulnerabilities. OpenAI rated it Critical under its Preparedness Framework, up from High for the previous version.

The release puts a frontier AI model into active vulnerability research at a scale human teams cannot match. It also tests whether the same capability can be kept from attackers, and whether maintainers can patch faster than the model finds bugs.

Why it matters

Software everywhere depends on maintainers finding bugs before attackers do. A model that finds them at this scale changes that race.

Questions about this story

Free account needed to ask — your question is kept and asked for you right after sign-up. Answers are public.

No questions yet — be the first to ask.

Key Indicators

48.5%

ExploitGym score

Share of specified vulnerabilities the model successfully exploited, up from 32.1% for GPT-5.6 Cyber.

99.93%

In-scope task adherence

Tests where the model avoided exploiting out-of-scope services without a deployment prompt.

New Firefox vulnerabilities

Previously unknown flaws Mozilla classified as high severity during a research campaign.

948

Linux kernel exploits

Working privilege escalation exploits the model developed, including 588 previously unknown vulnerabilities.

304

Critical Linux disclosures

Findings sent to Linux maintainers with candidate patches.

$1.2M

API cost of Linux run

Cost OpenAI reported for one autonomous vulnerability discovery and patching run.

Voices

Curated perspectives — historical figures and your fellow readers.

People Involved

Fouad Matin

Head of OpenAI's cyber preparedness and safety team

Leading safety evaluation of GPT-6 Cyber

Organizations Involved

OpenAI

Artificial intelligence company

Developer and deployer of GPT-6 Cyber

OpenAI is the San Francisco-based AI lab behind GPT-6 Cyber and the Daybreak access program.

Mozilla

Browser and software organization

Receiving and patching Firefox vulnerability reports

Mozilla maintains the Firefox web browser and classified 81 model-found flaws as high severity.

Daybreak

Access program

Tiered access channel for GPT-6 Cyber

Daybreak is OpenAI's tiered access program for security work, split into Blue and Red tracks.

Timeline

2 events Latest: Today

November 2026
OpenAI releases GPT-6 Cyber
Today Product launch

OpenAI releases GPT-6 Cyber to select Daybreak Red customers, rated Critical.
Today
Test run finds Firefox and Linux flaws
Today Research finding

Model's test run finds 81 Firefox flaws and 948 Linux exploits.
Today

Scenarios

Maintainers patch faster than GPT-6 Cyber finds new bugs

Possible Resolves by May 5, 2027

Discussed by: OpenAI researchers and Mozilla, based on the disclosed Firefox and Linux findings

Mozilla and Linux maintainers work through the disclosed reports and ship fixes. If the backlog clears and no new high-severity flaws from the model remain unpatched, the patch race tilts toward defenders.

Unauthorized actor misuses GPT-6 Cyber capabilities

Possible Resolves by Nov 5, 2027

Discussed by: OpenAI and security researchers tracking AI-enabled attacks

A group outside Daybreak Red obtains or replicates the model's exploit capability and uses it against a system without authorization. OpenAI says it is already seeing more misuse attempts and distillation efforts.

OpenAI expands GPT-6 Cyber to all approved Daybreak Red customers

Likely Resolves by Jan 5, 2027

Discussed by: OpenAI's launch statement

OpenAI said it will progressively expand access to all approved Daybreak Red customers in the coming weeks. The trigger is OpenAI's own rollout schedule, not an external event.

Regulators require pre-deployment review of Critical-rated cyber models

Unlikely Resolves by Nov 5, 2027

Discussed by: AI policy analysts and government AI offices

A regulator cites the Critical rating and the Firefox and Linux findings to require outside review before similar models are deployed. The trigger would be a formal rulemaking or binding guidance, not a statement of concern.

Historical Context

3 moments from history that rhyme with this story — and how they unfolded.

August 2016

DARPA Cyber Grand Challenge (2016)

Seven autonomous systems competed to find, exploit, and patch software flaws without human help. The winning system, Mayhem, ran at machine speed against a field of custom binaries.

Then

The contest showed that automated vulnerability discovery and patching could work under controlled conditions.

Now

It set a benchmark for machine-speed security research that commercial AI labs later pursued.

Why this matters now

GPT-6 Cyber is a commercial descendant of that idea, applied to real codebases like Firefox and the Linux kernel.

May 2017

EternalBlue and WannaCry (2017)

A leaked National Security Agency exploit called EternalBlue powered the WannaCry ransomware attack, which spread across hospitals, factories, and government networks in more than 150 countries.

Then

Hundreds of thousands of computers were affected before Microsoft's patch reached many systems.

Now

The attack became the standard example of how a powerful exploit can escape its intended scope.

Why this matters now

It shows why OpenAI's in-scope adherence numbers matter: a capable exploit tool that leaves its lane can cause global damage.

December 2021

Log4Shell (2021)

A single flaw in the widely used Log4j logging library let attackers run code on millions of servers. Security teams spent weeks patching systems across cloud providers and enterprises.

Then

Emergency patches rolled out through the holidays, and exploitation continued for months.

Now

The incident became a reference point for how one vulnerability can ripple through the software supply chain.

Why this matters now

GPT-6 Cyber's Linux findings show how many similar flaws may sit undiscovered, and how fast a model can surface them.