OpenAI's new cybersecurity model finds hundreds of bugs. Defenders still can't keep up.

Max Justicz, CTO and co-founder of Charge Robotics, pointed an OpenAI model at a chunk of Linux code and asked it to find bugs. Two hours later, he had a severe vulnerability on his hands.

Justicz has a hobbyist's interest in security research and had previously found a flaw in the same codebase. This time, using GPT 5.6 Sol through OpenAI's Daybreak Blue program (which gives vetted defenders access to frontier models), he essentially told the model to repeat history: "I literally just aimed ChatGPT at the same body of code and said 'I found this bug previously that was fixed several years ago, please find bugs as bad or worse,' and it came back in two hours with this." The result was CVE-2026-109518, a flaw that could let attackers take control of servers running Ubuntu, one of the most widely deployed Linux distributions on the planet.

He reported it to the Debian Project, which patched it. No evidence of exploitation has surfaced. But the speed and ease of the find is the real story here. Justicz called it "extremely convoluted and hard for a human to find," which is exactly the kind of bug that used to take a skilled researcher weeks, not a weekend.

"I can only imagine what one could do with more time and access to less neutered models," Justicz wrote in a social media thread about the discovery. "We should expect some extremely crazy bugs to be found and exploited over the next few months." His blunter point: "there is currently huge alpha in taking some code or binary someone has not audited yet and just aiming a powerful AI model at it."

OpenAI's answer: a dedicated cyber model, and a lot of scanning

OpenAI appears to have gotten the memo. The company announced ChatGPT Cyber, a model built specifically for cybersecurity work, rolling it out to the top tier of its Daybreak program, Daybreak Red, which is aimed at offensive security researchers doing vulnerability hunting and exploit development.

The timing isn't subtle. A day earlier, OpenAI said the same underlying model had been used to scan the Linux kernel and draft patches, surfacing 948 potential issues. It also found 81 "novel" vulnerabilities in Firefox. That's a lot of bugs, and a lot of signal that automated vulnerability discovery at scale is no longer theoretical.

• ChatGPT Cyber: a cybersecurity-focused model for Daybreak Red's offensive researchers
• Linux kernel scan: 948 potential vulnerabilities flagged, patches prepared
• Firefox scan: 81 novel vulnerabilities found
• Cost of the Linux hunt: $1.2 million in compute

OpenAI frames this as closing a narrow window. The company's pitch is that defenders need to move now, before AI-assisted hacking gets both more capable and more accessible to attackers. "Our goal is to dramatically expand Daybreak, bringing advanced security research to organizations of every size, independent researchers, and open-source maintainers," the company said in a statement.

That urgency cuts both ways, though. If frontier models are genuinely better at finding bugs than most human researchers, the same capability sitting in the wrong hands is a serious problem. OpenAI also announced a $1 billion initiative under Daybreak aimed at cybersecurity professionals protecting critical infrastructure like power grids and water systems, which tells you the company sees this as more than a product launch.

Eric Wallace, a researcher on OpenAI's technical staff, said the company is "now allocating a significant amount of our frontier compute towards discovering and patching vulnerabilities in open source codebases, fully pro bono." Good news for maintainers of underfunded open-source projects who've never had the resources for this kind of audit.

But the resource math is worth sitting with. The Linux bug hunt cost $1.2 million in compute. OpenAI reportedly spent roughly $6 million in tokens chasing a solution to a famous fluid dynamics problem. If patching the software that underpins the internet is less of a budget priority than a math puzzle, that's a signal about where the company's resourcing actually sits, intentions aside.

Justicz also flagged a practical gripe: he thinks OpenAI recently "nerfed" Daybreak Blue, making it less effective at finding vulnerabilities. There's a reasonable case for that: restricting raw capability limits what bad actors can extract if they get access, and OpenAI isn't exactly handing out invites casually. Justicz said he had to upload a photo of his driver's license just to join the program.

The flip side is that legitimate defenders are now working with one hand tied while attackers aren't playing by the same rules. Open-weight models with guardrails stripped out, sometimes called "abliterated" models, are already circulating and don't come with usage restrictions or ID verification. A defender capped by a nerfed model is still slower than an attacker running an uncensored one.

Scale is the other issue. OpenAI says "thousands" of cybersecurity professionals are currently in the Daybreak program. That's a rounding error against the global population of people capable of writing exploit code, especially once AI assistance lowers the skill floor for finding bugs in the first place. Justicz, for his part, said he personally spent about $5,000 on subscriptions and tokens just to find a handful of exploits, which isn't a trivial barrier for independent researchers or small open-source maintainers operating on no budget at all.

Key Takeaways

The core tension here isn't whether AI can find vulnerabilities fast. It clearly can, as Justicz's two-hour Ubuntu find and OpenAI's own kernel scan both demonstrate. The real question is whether the defensive side can scale access, funding, and researcher hours fast enough to match a capability that keeps getting better and keeps getting easier to access, nerfed or not. A $1.2 million kernel scan and a $1 billion defender fund are a start. Whether they're a big enough start is still an open question, and the clock OpenAI itself says is ticking doesn't leave much room for a slow answer.