OpenAI's GPT-6 Cyber finds and exploits software flaws, and goes first to trusted defenders

An AI built to break software, on purpose

OpenAI has released GPT-6 Cyber, a model aimed at security research. As Business Today reported, it is meant to spot weaknesses in software, check whether those weaknesses can really be abused by an attacker, and help fix them.

A vulnerability is a flaw in a program, a bit like an unlocked back door. An exploit is the method that uses that flaw to get in. Security teams already hunt for both, mostly by hand, and proving that a flaw is truly dangerous takes real skill and time.

GPT-6 Cyber is described as a major step up from GPT-5.6 Cyber, an earlier OpenAI model for the same kind of work. The new one is the version OpenAI is now putting in front of security professionals.

What the test scores say

OpenAI's main headline figure comes from ExploitGym, a benchmark, meaning a standard test used to compare models. GPT-6 Cyber scored 48.5% on it, against 32.1% for its predecessor. OpenAI's blog post, as relayed by Business Today, calls that a 51% improvement on exploit-specific vulnerabilities.

On another test, RemoteExploitBench, the model was pointed at business software including GitLab, PostgreSQL and Jenkins. It finished 95% of the tasks it was given.

There was also a harder setup. In black-box testing, where the model has no access to a program's source code (the underlying instructions), it had a 40.3% vulnerability recall rate. In plain terms, that is a measure of how many of the real flaws it managed to catch. OpenAI says this beat general-purpose models such as GPT-6 Astra.

These are the company's own figures from its own blog post, so they are worth treating as claims rather than independent results.

Real bugs in Firefox and Linux

The more concrete evidence is in actual software. OpenAI says the model is already being used to find and confirm thousands of exploits in important projects, including Linux and Mozilla Firefox.

• Researchers used it to uncover 81 Firefox flaws that nobody had known about before. Mozilla classed them as high-severity and has started shipping updates to fix them.
• For the Linux kernel, the core of the Linux operating system, it helped researchers build working exploits for 948 vulnerabilities. Of those, 588 were previously unknown.

The Firefox result is the clearest sign that this is not only a lab exercise. Flaws were found, a software maker agreed they were serious, and fixes are going out to users.

The risk side: a 'Critical' label and new guardrails

A tool that can prove how to break into software is a tool an attacker would also want. OpenAI appears to take that seriously. Because of what the model can do, the company moved its cybersecurity risk rating for GPT-6 Cyber up a level, from 'High' to 'Critical', under its internal safety framework.

OpenAI also reports a test called ExploitGym Honeypot, which checks whether the model keeps to the limits it has been given. Without extra safety prompts at deployment, it stayed in bounds 99.93% of the time. GPT-5.6 Cyber managed 94.49%.

The company lists several safeguards. These include ongoing monitoring for the model drifting away from its intended behavior, an approved-address list (called IP allowlisting) for a high-risk tool named Daybreak Red, and closer oversight where critical infrastructure is involved.

Who gets to use it

This is not a public release for everyone. Access is rolling out first to chosen cybersecurity partners, enterprise security teams and defenders OpenAI considers trustworthy. They reach it through Codex, OpenAI's coding tool, and through the API, which lets companies build the model into their own systems.

Business Today did not state a price, and nothing in the report says when access might widen.

The approach follows a common pattern for tools that cut both ways: put them first in the hands of people whose job is to defend, and watch how they are used before going further.