GPT 5.6 Sol Found a Critical Ubuntu Hole in Two Hours — and OpenAI Launches ChatGPT Cyber

Imagine this: you are a hobbyist security researcher. You take a codebase where a bug has already been found and tell the model: "find ones just as bad or worse." Two hours later you have a working vulnerability. That is exactly what Max Yustits, co-founder and CTO of Charge Robotics, did, and the result was unpleasant: the hole allows a server running Ubuntu to be taken over.

What happened

Yustits used GPT 5.6 Sol, released in July, through OpenAI's Daybreak Blue program. It provides access to powerful models for security defenders. The bug, he says, is convoluted and very hard for a human to find, but the exploit works. He reported the problem to the Debian project, which maintains the vulnerable software, and a patch has already been released. There are no signs of exploitation.

The cost is also telling: about $5,000 on subscriptions and tokens for several exploits found. For an attacker that is peanuts compared with the market price of a good 0-day. Yustits himself says there is currently huge "alpha potential" in taking unaudited code or a binary and pointing a powerful model at it.

OpenAI's response

The company announced ChatGPT Cyber, a specialized model for the top tier, Daybreak Red, where offensive researchers work. The day before, OpenAI said that the same model had scanned the Linux kernel and found 948 potential issues, and discovered 81 new vulnerabilities in Firefox. The Linux search cost $1.2 million. In addition, a $1 billion initiative was announced for specialists defending critical infrastructure: power grids and water supply.

OpenAI researcher Eric Wallace said the company directs a significant share of its frontier compute to finding and closing holes in open-source code for free. For comparison, an attempt to solve a well-known fluid dynamics problem consumed about $6 million in tokens, so the priorities are visible.

Where the weak spots are

The first problem is asymmetry. "Thousands" of defenders take part, while potential attackers are far more numerous. Yustits notes that Daybreak Blue was recently "nerfed" and became worse at finding vulnerabilities. Meanwhile, malicious actors can run open-weight models with the restrictions removed. Access to the program, incidentally, requires a photo of a driver's license, so identity verification does exist.

The second problem is the burden on maintainers. Frontier labs generate a mountain of findings, while there are not enough people to verify and patch them. Finding a bug has become cheaper than fixing it.

What to do in practice

• If you maintain open source or internal services, run your code through a strong model before someone else does. It won't be expensive: thousands of dollars, not millions.
• Pay special attention to old unaudited code and binaries.
• Apply to Daybreak if you are a defender or maintainer: OpenAI promises to expand the program.
• Speed up how you accept and roll out patches. The window between a vulnerability's publication and its mass exploitation is shrinking.
• Follow Ubuntu and Debian updates and install patches without delay.

The window of opportunity for defenders that OpenAI talks about is genuinely narrow. The question is not that models can search for bugs, but whether people will manage to close them in time. For now the answer is unclear, and it is safer to assume that someone is already doing the same thing with your code.