OpenAI released GPT-6 Cyber on November 5, a security model it rates Critical under its Preparedness Framework. OpenAI says it found 81 new high-severity Firefox bugs and built exploits for 948 Linux kernel flaws. Access starts with a few approved Daybreak Red customers through the API and Codex. No price was announced.
OpenAI has a new hacking model, and most companies can’t use it yet. GPT-6 Cyber, released November 5, finds software flaws and writes working exploits for them. OpenAI rates it “Critical” for cybersecurity under its Preparedness Framework, its highest risk tier, so it starts with a small group of approved customers in Daybreak Red, OpenAI’s program for offensive security work. For everyone else, the practical news is downstream: the bugs it already found are being fixed in software you run, so keep Firefox and your Linux kernels on automatic updates.
What OpenAI says it found
According to The Register and CyberScoop, OpenAI researchers used the model to:
- Find 81 previously unknown high-severity vulnerabilities in Firefox. The Register reports Mozilla has started fixing them.
- Develop working exploits for 948 Linux kernel vulnerabilities, 588 of them previously unknown. The Register says these are being reported with proposed patches for maintainers to review. CyberScoop reports OpenAI said all the flaws were reported and patched.
- Develop and validate a critical remote code execution flaw in Brotli, the open-source compression library, per CyberScoop.
These numbers come from OpenAI. Mozilla and the kernel maintainers had not published their own counts at publication.
Why OpenAI is limiting access
OpenAI’s Critical tier means a model can find and build working zero-day exploits in real systems and run attacks end to end, according to The Register. That is the reason for the gate. GPT-6 Cyber is available through the API and Codex only to selected Daybreak Red customers, who are approved for exploit development, penetration testing and red teaming. OpenAI says it will widen access to all Daybreak Red members over the coming weeks and is also opening GPT-5.6 Cyber to that group.
The release leans on behavior as much as skill. OpenAI says the model is “trained to stay within authorized scope.” On its ExploitGym benchmark it scored 48.5%, against 32.1% for GPT-5.6 Cyber, and in a separate test it avoided exploiting out-of-scope services more than 99% of the time, per The Register. OpenAI declined to tell The Register what happened in the 0.07% of cases where the model went past its intended scope. CyberScoop reports new safeguards: stricter identity and account security, more frequent check-ins with OpenAI before high-risk actions, and extra monitoring for rule-breaking.
What security teams should check
If your company is in OpenAI’s Daybreak Blue program for defensive work, read the data terms. CyberScoop reports OpenAI told members last week that, from the end of the month, Daybreak Blue sessions on personal accounts would be used to improve its models, then clarified that customers can keep opting out. Confirm your account’s setting in writing before you put client code through it.
If you run a small team without a red team, nothing here is for sale to you today. The useful move is the boring one: patch windows matter more when AI finds bugs by the hundred. Volta, the infostealer that uses AI to rank its victims, shows attackers are moving just as fast. And this week’s OpenSSH *sshole fix and the Linux kernel TCP flaw are reminders that patch notes now arrive faster than most teams read them.