OpenAI strengthens GPT-6 Cyber for vulnerability detection
OpenAI unveils GPT-6 Cyber, a variant of GPT-6 Astra reserved for approved clients. The model targets vulnerability research, with notable results on Firefox and Linux.
On November 5, 2026, OpenAI introduced GPT-6 Cyber, a cybersecurity-specialized model derived from GPT-6 Astra and reserved for a select group of approved clients. The tool aims to help researchers and defense teams detect, reproduce, and validate software vulnerabilities faster. In a sector where the time between discovery and patch matters as much as the flaw itself, the goal is clear: reduce exposure time. And prevent bugs from escaping before security teams.
A model designed for security teams, not for the general public
GPT-6 Cyber does not present itself as just another general-purpose assistant. OpenAI explains that it deeply analyzes code, reproduces malfunctions, develops proof-of-concepts to verify exploitability of a flaw, and can propose fixes. In other words, it fits into the workflow of security researchers, not general public use. The company aims to shorten the period between vulnerability discovery and its correction by the concerned vendor.
The model is classified at the "critical" level in cybersecurity, a first for OpenAI's Cyber models. The logical consequence: it is not freely accessible. Access is first granted via the API and Codex, to a limited group of approved clients in the Daybreak Red program. OpenAI plans a gradual opening to other eligible organizations. Meanwhile, Daybreak Blue, the offer for defense teams, also gains access to advanced models for vulnerability research, patch validation, and strengthening security operations.
81 new vulnerabilities in Firefox, and a rising score
Initial figures provided by OpenAI give an idea of the model's positioning. In tests conducted on Firefox, GPT-6 Cyber identified 81 previously unknown vulnerabilities, classified as high severity by Mozilla. The model also uncovered 30 known issues whose severity had been underestimated. These findings were shared with Mozilla along with reproduction details and patch proposals, and the vendor has already integrated several fixes.
OpenAI also cites work on the Linux kernel. The model reportedly reproduced 1,724 vulnerabilities and developed 948 functional exploits. These results do not fully reflect the real quality of a security tool, but they show that AI can already automate part of the vulnerability identification and validation work. This is where the machine stops being a mere co-pilot and becomes a colleague that’s a bit too fast.
What benchmarks show
On ExploitGym, a benchmark evaluating a model's ability to exploit a known vulnerability in a controlled environment, GPT-6 Cyber reaches 48.5%, compared to 32.1% for GPT-5.6 Cyber. OpenAI also reports a detection rate of 40.3% during evaluations on web applications without source code access, versus 32% for GPT-6 Astra. These scores should be interpreted cautiously, as they describe test conditions, not a real production environment.
For cybersecurity actors, the issue goes beyond raw performance. If AI helps find exploitable flaws faster, it can also accelerate patch cycles on the vendor side and strengthen defense. The question of governance remains. OpenAI says it has reinforced controls around the API, with identity verification, multi-factor authentication, IP address filtering, and monitoring mechanisms for out-of-scope usage. In this field, the tightrope is narrow: what helps defend can also be used to attack.
Key points
- GPT-6 Cyber was presented on November 5, 2026.
- OpenAI reserves it for approved clients via Daybreak Red.
- Mozilla received 81 new vulnerabilities on Firefox.
- The model scores 48.5% on ExploitGym.
- Daybreak Blue targets defense teams.
- OpenAI classifies the model at the "critical" level.
By the numbers
- 81 vulnerabilities — new, detected on Firefox, according to OpenAI and Mozilla.
- 30 issues — already known, but whose severity was underestimated.
- 1,724 vulnerabilities — reproduced on the Linux kernel, according to OpenAI.
- 948 functional exploits — developed from work on Linux.
- 48.5% — GPT-6 Cyber's score on ExploitGym, versus 32.1% for GPT-5.6 Cyber.