OpenAI's GPT-6 Cyber built 948 working Linux kernel exploits at about $1,279 each. Audit your clients' custom code now

Finding exploitable bugs just got a lot cheaper, and the glue code in your client automations is code too.

What happened

OpenAI introduced GPT-6 Cyber, a model trained to find software vulnerabilities, build working exploits and suggest patches. According to OpenAI, as reported by The Decoder, a targeted campaign found 81 previously unknown severe vulnerabilities in Firefox, which Mozilla has already fixed. In the Linux kernel, an automated system of scout and analyst agents reproduced 1,724 vulnerabilities and built working exploits for 948 of them, more than five times the 166 that GPT-5.6 Sol managed under the same conditions. OpenAI puts the compute cost of that run at $1,213,430, about $1,279 per exploit. The Decoder compares that with bounty programs that have historically paid $20,000 to $70,000 per comparable finding. All benchmark numbers come from OpenAI, with no independent testing published yet.

Access is limited for now to select customers in OpenAI's Daybreak Red program, through the API and Codex. OpenAI recommends running the model in isolated environments with human oversight.

TechCrunch adds a smaller story that matters more for most of us. Max Justicz, CTO of Charge Robotics, used an earlier model through OpenAI's Daybreak Blue program to find a severe Ubuntu vulnerability. He says it came back in two hours. He reported it to the Debian Project, which issued a patch. He told TechCrunch he spent about $5,000 on subscriptions and tokens to find several exploits.

My take

The headline is a frontier model. The lesson for operators is the price curve. If one person with a few thousand dollars can aim a model at code nobody has audited, then the custom code we ship for clients is now cheap to probe.

Think about what sits in a typical automation stack: a self-hosted n8n instance, webhook endpoints that accept any payload, small scripts that parse form data, API keys in environment variables. None of it gets the scrutiny of a browser or a kernel.

What I would do this month:

1. List every public endpoint each client automation exposes.
2. Require a secret or signature on every inbound webhook.
3. Run a model you are allowed to use over your own custom code nodes and scripts, then fix what it finds.
4. Rotate any key that has lived in a workflow for more than a year.

Defenders get these tools in stages. Assume attackers do not wait.