OpenAI is giving its top-tier security researchers a new model built for vulnerability hunting, as evidence mounts that AI can find serious software bugs at low cost. TechCrunch AI reports that ChatGPT Cyber is rolling out to Daybreak Red, the highest tier of OpenAI’s cybersecurity program, which is aimed at offensive security researchers. This is real progress, but the numbers suggest the effort still doesn’t match the size of the problem.
The short version
- The launch: ChatGPT Cyber goes to Daybreak Red users for finding and developing vulnerabilities.
- The results so far: OpenAI says the model flagged 948 potential issues in the Linux kernel and prepared patches. OpenAI also reports 81 “novel” vulnerabilities in Firefox.
- The money: OpenAI announced a $1 billion initiative for program members who protect critical infrastructure such as electrical grids and water systems.
- The catch: OpenAI says only “thousands” of security professionals are in Daybreak, and the potential pool of hackers is far larger.
A two-hour prompt found a severe bug
The clearest warning came from outside OpenAI. Max Justicz, CTO and co-founder of Charge Robotics and a self-described hobbyist in security, used GPT 5.6 Sol through the lower-tier Daybreak Blue program. He found a severe vulnerability that could let attackers take control of servers running Ubuntu.
His method was almost comically simple. “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,” Justicz told TechCrunch.
He reported the flaw to the Debian Project, which maintains the affected software and issued a patch. There’s no known evidence that anyone exploited it. Justicz says he spent about $5,000 on subscriptions and tokens to find several exploits, which is a very low bar for this kind of work.
His read on what comes next is blunt: “We should expect some extremely crazy bugs to be found and exploited over the next few months.”
Why this matters
Bug hunting used to be limited by how many skilled people could spend weeks reading difficult code. That limit is weakening fast. Justicz described his bug as “extremely convoluted and hard for a human to find,” yet a model found it in an afternoon.
That creates an awkward split. Finding bugs is getting cheap, but fixing them still takes people, testing, and release cycles. Many defenders feel the frontier labs have handed them a mountain of new work without the resources to clear it, and open-source maintainers are in a tough spot if reports pile up faster than they can review them.
OpenAI’s guardrail problem
Daybreak is also caught between two bad options:
- Too loose: Justicz says he joined the program by uploading a photo of his driver’s license. That’s a thin check for access to tools this strong.
- Too tight: Justicz also says OpenAI appears to have recently “nerfed” Daybreak Blue, making it less useful for finding vulnerabilities. Security professionals argue that this leaves them with one hand tied, because attackers can run open-weight models with “abliterated” guardrails, meaning the safety limits have been stripped out.
So restricting frontier models only slows attackers who rely on those models. It doesn’t touch anyone running open weights on their own hardware.
The number I’d watch
OpenAI researcher Eric Wallace says the company is “allocating a significant amount of our frontier compute towards discovering and patching vulnerabilities in open source codebases, fully pro bono.”
The receipts make that claim look less impressive. The Linux bug hunt cost $1.2 million, about a fifth of the roughly $6 million in tokens OpenAI spent chasing a solution to a famous fluid dynamics problem. By my math, that’s around $1,270 per potential kernel issue. OpenAI keeps saying defenders have a narrow window before these models get better at hacking and easier to access. If that’s true, its Linux spending looks small for a kernel that underpins so much of the world’s infrastructure.
What to do now
- Audit your own code first: Justicz says there’s “huge alpha” in pointing a strong model at code nobody has audited. Attackers can do the same thing, so assume your unaudited code will get scanned.
- Shorten patch cycles: The gap between disclosure and patching gets riskier when bugs are this cheap to find.
- Plan for triage: The 948 figure covers potential issues, not confirmed ones. Someone still has to check every report, so budget for that time.
OpenAI says it wants to expand Daybreak to organizations of every size, independent researchers, and open-source maintainers. The test is whether that happens before the “extremely crazy bugs” Justicz expects start showing up in the wild. The full story is available at TechCrunch AI.