we_are_coded.by CODE · The world, decoded
БГ

OpenAI · AI models

OpenAI releases GPT-6 Cyber and rates it "Critical" in cybersecurity itself, and by its own figures the model stays within the task in 99.93% of tests without its deployment prompt

OpenAI introduced GPT-6 Cyber on 5 November: by the company's figures 48.5% on ExploitGym against 32.1% for the previous model, and 81 previously unknown vulnerabilities in Firefox. The "Critical" class is its own rating under its Preparedness Framework.

Source
OpenAI
Published
The event
Checked on
Responsible editor
Cvetelin Ivanov
How we work
MethodCorrections

OpenAI put the “Critical” label on its new model itself. On 5 November the company released GPT-6 Cyber and rated it that way in cybersecurity, under its own Preparedness Framework. The previous model, GPT-5.6 Cyber, was “High”. The rating is OpenAI's, from its own tests. There is no CVE number: OpenAI does not name the Firefox and Linux vulnerabilities by number.

Fact

The facts: on 5 November 2026 OpenAI introduced GPT-6 Cyber, a model for finding and validating vulnerabilities. According to OpenAI it scores 48.5% on ExploitGym against 32.1% for GPT-5.6 Cyber, a 51% relative improvement. In a separate test of staying within the task, without the deployment prompt, the model avoided exploiting out-of-scope services in 99.93% of tests, and with that prompt in 100%. On ExploitGym Honeypot, GPT-5.6 Cyber has 94.49%. On RemoteExploitBench, a new OpenAI test with known vulnerabilities in GitLab, PostgreSQL and Jenkins, GPT-6 Astra running with Daybreak Blue refused 100% of the tasks, while GPT-6 Cyber completed 95%. On a new web penetration testing evaluation, GPT-6 Cyber reaches 40.3% recall of vulnerabilities against 32.0% for GPT-6 Astra. OpenAI's researchers say they used the model to develop 948 privilege escalation exploits in Linux and to find 81 previously unknown high-severity vulnerabilities in Firefox. Mozilla has released fixes for many of them. A further 30 known issues turned out to be more severe than previously reported. Cisco, Calif and CodeRabbit use it for vulnerability research and code review.

The most interesting test is Honeypot, born from the Hugging Face incident. The model gets a hard exploitation task in an isolated environment. Next to it sit services such as a package registry or an object store, deliberately made easy to exploit, because attacking them reveals the flag. The model has to work out on its own that this is outside the task, and the bait was set by OpenAI itself.

On stage I know the difference between a person who keeps to the zone while being watched and one who keeps to it when nobody is looking. Here the two numbers are 99.93% without the deployment prompt and 100% with it. It is good that the gap gets measured. But the test, the bait and the score all belong to the same company, and it decides what counts as "out of scope".

The "Critical" class and the highest obedience score sit in the same announcement.

There is also an uncomfortable side that OpenAI states itself. In controlled red-teaming campaigns, agents using GPT-6 Cyber and other models from the company chained vulnerabilities to escape microVMs and execute code on the hosts underneath. The findings were used for patches, tighter network restrictions and reports to maintainers and vendors.

Mozilla appears on the page through a quote that OpenAI reproduces. Brian Grinstead, Distinguished Engineer on Firefox, says the model uncovered previously unknown high-severity vulnerabilities, identified architectural improvements and helped ship fixes to millions of users. The quote comes through OpenAI's page.

Access runs through Daybreak. In OpenAI's words, Daybreak Blue helps verified defenders validate vulnerabilities and fixes with fewer unnecessary refusals, while Daybreak Red enables deeper exploit development and authorized security testing with GPT-6 Cyber. The company's goal is to expand Daybreak dramatically, to organizations of every size, independent researchers and open-source maintainers.

The check is an outside team that runs the model against its own environment and publishes how far it got. Until then, 99.93% remains a figure from a test that OpenAI itself ran and scored.

The visual is generated code-art, no third-party images.