OpenAI released a new cybersecurity model. Is that enough?

OpenAI announced ChatGPT Cyber, a model aimed specifically at cybersecurity work, and is making it available to the highest tier of its Daybreak Red program, which gives offensive security researchers tools for hunting and developing vulnerabilities. The announcement came the day after OpenAI said the same model had been used to scan the Linux kernel and prepare patches for what it found, turning up 948 potential issues, and that 81 "novel" vulnerabilities had been identified in the Firefox browser.
The company framed the release as a step toward widening access rather than narrowing it. In a statement it said the goal is to dramatically expand Daybreak, bringing advanced security research to organizations of every size, independent researchers and open-source maintainers. Alongside model access, OpenAI has announced a $1 billion initiative to support cybersecurity professionals in the program who work on critical infrastructure such as the electrical grid and water systems.
Participation is not open. OpenAI says only "thousands" of cybersecurity professionals are currently in the Daybreak program, a small figure against the far larger pool of potential attackers the company cites. Vetting appears light at the entry level: Max Justicz, chief technology officer and co-founder of the startup Charge Robotics, said he joined by uploading a photo of his driver's license.
The practical capability is already visible outside OpenAI's own announcements. Justicz, who has a hobbyist interest in security and had previously found a bug in the same code, said he used GPT 5.6 Sol — released in July — through Daybreak Blue to find a severe vulnerability that, if exploited, could let attackers take control of servers running Ubuntu, the widely used open-source Linux distribution. He described aiming the model at the codebase and asking it to find bugs as bad as or worse than one he had found and that was fixed several years ago; the model returned the result in about two hours. He called the flaw extremely convoluted and hard for a human to find, but functional. He reported it to the Debian Project, which maintains the software in question and issued a patch. There is no known evidence the flaw was exploited.
Justicz's read is that this is a preview rather than a peak. He wrote in a social media thread that he could only imagine what someone could do with more time and access to less restricted models, and that some extremely serious bugs should be expected to be found and exploited over the coming months. He added that there is currently huge opportunity in taking code or binaries nobody has audited and pointing a powerful model at them. That is a fair summary of the economics in play: a model that can read a large codebase cheaply changes the cost of finding certain classes of memory-safety and parsing bugs, and defenders are not the only ones who can pay that cost.
Cost, though, is not trivial for individuals. Justicz said he personally spent roughly $5,000 on subscriptions and tokens to find several exploits. At the organizational level the numbers are larger. OpenAI researcher Eric Wallace said the company is now allocating a significant amount of its frontier compute toward discovering and patching vulnerabilities in open-source codebases, fully pro bono. The Linux bug hunt cost $1.2 million, which the report notes is well below the roughly $6 million in token spend OpenAI put toward finding a potential solution to a famous fluid dynamics problem.
Daybreak itself is drawing criticism from the researchers it is meant to serve. Justicz said OpenAI appears to have recently "nerfed" Daybreak Blue, making it less useful for finding vulnerabilities. That cuts both ways. Limiting capability reduces what a bad actor could do after signing up with little more than an ID photo, but security professionals argue they are working with one hand tied if frontier tools are throttled for them while attackers can run open-weight models with safety guardrails stripped out — the practice sometimes called abliteration, where refusal behavior is removed from a model's weights. This is the central tension in the release: the same capability that makes a defender effective makes an attacker effective, and the tiering, vetting and tuning decisions are where OpenAI is drawing the line.
OpenAI's stated position is that there is a narrow window for the people who safeguard digital infrastructure to act before models get better at hacking and easier to obtain. For defenders, the experience can feel like the opposite: frontier labs have generated a volume of new findings that outstrips the staff available to triage and patch them. A count of 948 potential kernel issues is impressive as a demonstration and daunting as a work queue, since potential is not the same as confirmed, and each one needs human review before a patch ships.
The broader point is that vulnerability discovery has become a compute expense. Tasks that once required scarce expert time — auditing an unfamiliar codebase, drafting a candidate fix, narrowing a crash to a root cause — can now be bought with tokens, at least for some bug classes. That shifts the bottleneck from finding bugs toward verifying, patching and distributing fixes, which is slower, more political and harder to scale than model inference. It also raises a disclosure question the security community has wrestled with for years: when a model can surface hundreds of issues in widely deployed software, coordinated disclosure timelines built around individual researchers may not keep pace.
Why it matters: Developers maintaining open-source projects and teams running Ubuntu or other Debian-derived systems are on the receiving end of this, whether or not they ever touch a frontier model — patches will arrive faster and in greater number, and triage capacity becomes the constraint (an inference, not something OpenAI stated). For product teams, the practical question is whether to budget for AI-assisted security review, which Justicz's roughly $5,000 personal spend suggests is within reach of small companies, or to wait for lab-funded programs aimed at critical infrastructure. The tiering of Daybreak Red and the reported throttling of Daybreak Blue also mean the most capable tools may stay gated behind vetting, leaving smaller maintainers dependent on the goodwill of labs that are still weighing how much access to grant.
Based on reporting from the original publisher. Visit the source for full context and later updates.
Publisher excerpt
OpenAI's newest cybersecurity model comes out as AI labs face scrutiny of their work to protect digital infrastructure.