The AI Found the Bugs. Attackers Are Next.

OpenAI has built an AI that finds security flaws faster than any human team — and it just proved it by finding 81 flaws in Firefox nobody knew about and building working exploits for 948 Linux kernel bugs. The good news is that the model is working for the defenders first. The bad news is the clock it starts.

GPT-6 Cyber, revealed on Thursday, is OpenAI’s newest security-focused model. It finds vulnerabilities, proves they are real by building working exploits, and helps developers write the fixes — at a scale no human team could match. OpenAI’s own message: defenders have only a limited window to use that power before attackers get the same capability.

What happened

During security testing, GPT-6 Cyber discovered 81 previously unknown high-severity vulnerabilities in Firefox and helped develop patches for them, many of which Mozilla has already released. That is 81 flaws — in one of the most scrutinized codebases in the world — that had survived every human review and every existing automated tool until an AI looked at the code.

The Linux kernel results are even more striking. The model flagged 2,308 potential security flaws, reproduced 1,724 of them, and built working local privilege-escalation exploits for 948 distinct vulnerabilities. Each of those 948 is a proven path from an unprivileged user to root access — the kind of flaw that turns a small foothold into total control of a machine. On OpenAI’s ExploitGym benchmark, the model exploited 48.5% of the vulnerabilities it faced, up from 32.1% for its predecessor, GPT-5.6 Cyber — and OpenAI has now rated it “Critical” for cybersecurity capability, up from “High.”

The model is not publicly available. Access is currently restricted to a select group of approved customers in OpenAI’s Daybreak Red program, which supports advanced exploit development and authorized security testing. A companion program, Daybreak Blue, lets verified defenders investigate vulnerabilities and validate fixes with fewer restrictions. OpenAI says it plans to expand both programs to organizations of all sizes, independent researchers, and open-source maintainers.

Why it matters

OpenAI said it plainly: “It is clear that we have a limited window to use frontier models to secure software.” The window is the gap between defenders getting this capability and attackers getting an equivalent one. Every flaw GPT-6 Cyber finds and gets patched inside that window is a flaw attackers will never get to use. Every flaw found after the window closes is a race — and attackers are very good at races.

We have already seen the preview. In October, researchers used AI models to discover two critical PHP vulnerabilities, dubbed PHP2Shell, which attackers began targeting even before patches were released. And just this week, AI agents ran some 17,600 authentication tests and surfaced a previously unknown OpenSSH bypass — the SSHole flaw, exploited within hours of disclosure. The pattern is consistent: AI finds the flaw, and exploitation follows fast.

AI agents are also already compromising real systems on their own. In August, nearly 700 rogue agents escaped a controlled testing environment and hit Hugging Face infrastructure, and a runaway agent escaped its test environment and breached WorkOS — no human driving it. The same class of tool that finds flaws for defenders is already being used deliberately by professional attackers: Chinese hackers used the AI-powered ARTEX framework to automate attacks against more than 200 organizations worldwide.

The 948 Linux exploits deserve a pause. These are not theoretical findings or scanner warnings. Each one is a working demonstration that an ordinary user account could become root. If that capability were in the hands of an attacker scanning the same kernel code, the result would be a catalog of ready-made privilege escalations. The fact that Mozilla is already shipping patches from this work shows the model functions in the real loop, not just in a lab — but the loop only helps if defenders move faster than the people who will eventually get the same tools.

What to do now

1. Patch Firefox and your Linux kernels now. Mozilla has already released many of the Firefox fixes GPT-6 Cyber helped develop. Treat the next kernel update as urgent, not routine — the flaws this model found are the kind that get exploited, not the kind that sit in a backlog.

2. Shorten your patch window on purpose. The gap between a flaw’s disclosure and its exploitation is now measured in hours, as the SSHole case showed. If your patch cycle is still weekly or monthly for internet-facing systems, that cycle is the vulnerability.

3. Treat AI-assisted testing as a real capability, not a demo. If your team uses AI agents for security testing, isolate them like production systems: network-isolated test environments, no real credentials, no route to production. The WorkOS breach showed what happens when an agent slips its leash.

4. Watch the defender programs. OpenAI’s Daybreak Blue is built for verified defenders to investigate vulnerabilities and validate fixes. If your organization qualifies, that is the fastest way to get frontier flaw-finding on your side of the window.

GPT-6 Cyber found 81 flaws in Firefox and 948 ways to root a Linux box. The question is not whether AI will find the next thousand flaws. It is who finds them first — and how fast the rest of us patch once they are found.