OpenAI Releases GPT-6 Cyber, a Model That Finds Vulnerabilities and Proposes Fixes
The model scored 16.4 percentage points higher than its predecessor on a benchmark that tests whether it can actually exploit vulnerabilities. It is designed to help approved security professionals and companies conduct security testing, and was separately trained to stay within authorized boundaries.
OpenAI unveiled GPT-6 Cyber on November 5 local time. The model hunts down security vulnerabilities and proposes fixes for them. It also writes code that verifies whether a weakness in a piece of software can actually be exploited. OpenAI said it pushed the model’s offensive capability higher while separately training it to stay within the targets and the task boundaries that researchers have authorized.
How much better is it at security testing?
On ExploitGym, a benchmark that measures whether a model can actually exploit a designated vulnerability, GPT-6 Cyber scored 48.5%. That is 16.4 percentage points above the 32.1% posted by its predecessor, GPT-5.6 Cyber. On a test that requires finding vulnerabilities in web applications without access to the source code, it surfaced 40.3% of all vulnerabilities present, ahead of GPT-6 Astra’s 32.0%.
The model was also put to work finding vulnerabilities and confirming exploitability in software that is in real-world use. According to OpenAI, GPT-6 Cyber found 81 new vulnerabilities in Firefox and produced the tests and supporting evidence needed to reproduce them. Mozilla classified those vulnerabilities as high severity. OpenAI passed along proposed fixes as well, and Mozilla has already shipped patches for a substantial share of the vulnerabilities after receiving the reports.
In the Linux kernel, the model reproduced 1,724 vulnerabilities and, for 948 of them, went on to produce working exploit code. That is a sharp increase over the 166 achieved by GPT-5.6 Sol using the same research tooling. OpenAI said it forwarded 304 serious findings to maintainers along with proposed fixes. In other words, the work moved beyond simply spotting flaws to assessing the real-world damage potential and the remedy at the same time.
Does it only attack authorized targets?
Whether the model respects its authorized scope was evaluated separately. OpenAI ran tests in which auxiliary services unrelated to the assigned task were deliberately made easy to attack. In that setup, GPT-6 Cyber avoided out-of-scope attacks 99.93% of the time even without being given any separate safety instructions, the company said. GPT-5.6 Cyber managed 94.49%. With safety instructions applied, scope compliance reached 100%.
OpenAI raised the model’s cybersecurity risk in its internal risk framework from the previous model’s “High” to “Critical.” As safeguards against misuse, it said it monitors whether the model’s actual behavior lines up with the work the user authorized, and that it has strengthened screening designed to filter out dangerous activity. When developers plug Daybreak Red, which is meant for advanced security testing, into their own services, access is as a rule restricted to pre-approved networks.
Who can use it at work?
According to reporting by AI Times, the service has launched for a limited set of approved security professionals and companies. There are two routes: developers wiring it into their own services, and use through Codex. OpenAI describes Daybreak Blue as a tool for verified defenders to review vulnerabilities and fixes, and Daybreak Red as a tool for generating exploit code with GPT-6 Cyber and running authorized security tests. A separate access track for approved security personnel at critical infrastructure organizations was announced alongside them.
With this release, the range of AI assistance available to security teams widens from finding vulnerabilities to confirming exploitability and drafting fixes. If Korean companies weigh adoption, the deciding factors will likely be how the model performs on real-world security testing and how reliably it stays inside authorized boundaries. The next things to watch are whether the vulnerabilities it finds actually get fixed, and what results the tool delivers in the field.