On November 5, OpenAI unveiled GPT-6 Cyber, a version of its GPT-6 Astra model entirely dedicated to cybersecurity research. In just a few weeks of testing, this artificial intelligence helped detect dozens of critical vulnerabilities in software used by hundreds of millions of people, including the Firefox browser. It has reignited the debate over AI's growing role in both digital defense and digital attack.
Critical Flaws Found by an AI
According to OpenAI, GPT-6 Cyber helped identify 81 previously unknown vulnerabilities in Firefox, rated high-severity by Mozilla, along with roughly thirty already-known flaws whose severity had been underestimated. The browser maker has since released fixes for most of them. On the Linux kernel, a chain of agents driven by this model reproduced more than 1,700 vulnerabilities and developed 948 working privilege-escalation exploits, within a strictly controlled research setting.
These results illustrate a broader trend: large language models are increasingly able to analyze code in depth, reproduce malfunctions, and build proof-of-concept exploits to verify that a flaw is actually exploitable—work that until now required entire teams of specialized researchers.
Deliberately Restricted Access
Aware of the potential for this kind of tool to fall into the wrong hands, OpenAI classified GPT-6 Cyber at the “Critical” level of its preparedness framework, a first for a model in its Cyber lineup. Access is not open to the general public: it is reserved for a limited number of verified customers under the “Daybreak Red” program, with enhanced security requirements such as multi-factor authentication and authorized IP address filtering. Companies including Mozilla, Cisco, and Palo Alto Networks were among the teams able to use the model before its public announcement.
When used without framing instructions, the model stays within its authorized scope 99.93% of the time, compared with 94.49% for the previous generation. With appropriate deployment guidelines, OpenAI says it achieves 100% compliance with the set boundaries.
Why This Matters for the General Public
For the average internet user, the stakes aren't immediately visible but are very real: every flaw fixed thanks to this kind of tool reduces the risk of a hacker exploiting it before a vendor like Mozilla can respond. It's also a strong signal to the industry: vulnerability research, long reserved for highly specialized experts, is becoming accessible to AI agents capable of working at scale. That could speed up the patching of flaws, but it could also, if such tools were to leak or be copied without the same safeguards, multiply malicious uses.
For more on related topics, check out our article on Sam Altman and Dario Amodei's call for international AI rules, or our guide to choosing a truly effective antivirus. The subject also connects to advances in cloud-based quantum computing, another area where AI and cutting-edge computing meet.
Find more news in our Tech & Web section and our Science features.
Frequently Asked Questions
Is GPT-6 Cyber Available to the Public?
No. It is reserved for a limited number of verified customers in OpenAI's “Daybreak Red” program, with enhanced security requirements such as multi-factor authentication.
Have the Flaws Found in Firefox Been Fixed?
Mozilla has released fixes for most of the 81 high-severity vulnerabilities identified with the help of GPT-6 Cyber.
Why Does OpenAI Classify This Model as “Critical”?
Because its ability to find and exploit software flaws could, in the wrong hands, be used for large-scale cyberattacks: this is the first time a model in the Cyber lineup has received this classification level.
Sources
- OpenAI — Introducing GPT-6 Cyber
- Le Monde Informatique — OpenAI dévoile GPT-6 Cyber
- The Register — OpenAI releases GPT-6 Cyber
This article was written with the help of artificial intelligence. Editorial policy













