OpenAI's model built 948 working Linux exploits in one run

The demonstration

On 5 November 2026, OpenAI demonstrated that its upcoming AI model had automatically generated 948 working root exploits for the latest Linux kernel in a single run ^[1]. Each exploit worked. Each came with a corresponding patch. The model didn't just find theoretical problems; it turned them into working attack code, at a scale no human team could match in the same time.

It's a bit like handing a locksmith a street full of front doors and being told an hour later which 948 of them can be opened with a paperclip, and here's how. The point isn't whether the locksmith is good or bad. The point is the speed.

What it means for cybersecurity

AI has been used to find vulnerabilities for years, but this demonstration is different in two ways. First, it's the scale: nearly a thousand working exploits in one go, not a handful. Second, it's the automation: the model did the work without someone telling it where to look.

For defenders, that changes the timeline. If an AI can generate exploits this quickly, the window between discovery and active exploitation shrinks. For organisations selling or running software with digital elements, it raises a harder question: when your own AI, or someone else's, finds a problem in your product, how do you prove when you knew and what you did about it?

The reporting duty under CRA

The EU's Cyber Resilience Act has applied since 11 September 2026 ^[2]. Article 14 requires manufacturers of products with digital elements to report actively exploited vulnerabilities and severe incidents through ENISA's common platform, in stages with tight deadlines ^[2].

The first stage is an early warning, due within 24 hours of becoming aware. The clock starts when the manufacturer knows, not when it's convenient. If an AI model generates 948 exploits overnight and someone notices in the morning, those 24 hours have already begun.

The penalty for failing to report can reach €15 million or 2.5 per cent of global annual turnover, whichever is higher ^[2].

The evidence question

Say a supervisor asks why an early warning went out after 30 hours instead of 24. The manufacturer says the AI flagged the vulnerability on Tuesday morning, and the team escalated it immediately. The log says the same. But logs are kept by the people who could edit them, and a month later it's hard to prove the timeline is exactly what was recorded on the day.

That's the evidence question. It isn't whether the AI found something—it did. It's whether the organisation can show, in a way that holds up under scrutiny, when each step happened: detection, awareness, escalation, decision and report.

Screenshots can be altered. Logs can be backdated. When the people keeping the records are the same people whose compliance is being questioned, trust gets thinner.

How a digital witness fits here

This is a case where a digital witness fits well.

What you seal is each step of the incident timeline as it happens: the first detection note when the AI flags the vulnerability, the decision to escalate it, and each report as it is sent to ENISA's platform. Sealing takes a fingerprint of the record, not the record itself, so the detection note and the report never leave your systems.

You seal at each step. When the AI writes the detection note, you seal it. When someone decides to escalate, you seal that decision. When the early warning goes out, you seal a copy of what was sent.

Internal audit, supervisors and incident responders can check it. Anyone you share the record with can calculate the fingerprint again and compare it with the sealed one. If the record hasn't changed, the fingerprints match. Change a single character and they won't.

What it shows is when each step was recorded, in a way nobody can quietly rewrite afterwards, even someone with administrator rights. What it doesn't show is that the report was complete, that the response was adequate or that the organisation met every other duty under the regulation. It answers one question: did this record exist at this time, unchanged since?

How it would work

1. The AI flags a vulnerability and writes a detection note. Your system sends a fingerprint of that note to WitniumChain through the REST API, SDK or MCP server and gets a witness ID back.
2. The security team decides to escalate. That decision is sealed the same way.
3. The early warning is sent to ENISA. A copy of what was sent is sealed at the moment it goes out.

Months later, if the timeline is questioned, you share the records and the witness IDs. Anyone can check the fingerprints and see that the records haven't changed since they were sealed. The content stays with you. The first 2,000 witnesses a month are free.

What this means for you

If you manufacture software or hardware with digital elements, or if you run AI models that find vulnerabilities, here are three questions worth asking now:

• When an AI flags a vulnerability, how do you record when that happened and who saw it first?
• Could you show a supervisor or an auditor that your incident timeline is exactly what was recorded on the day, not edited later?
• If the 24-hour clock for an early warning starts when the AI detection note is written, does your system timestamp that moment in a way that can be independently checked?

The CRA's deadlines are short, and they start from awareness, not from convenience. A checkable timeline, sealed as it happens, turns a long argument into a short one.

What to watch

The Cyber Resilience Act's basic security requirements, including logging and vulnerability handling, apply from 11 December 2027 to products placed on the market after that date ^[2]. For manufacturers already covered by the reporting duty, the timeline requirements are live now.

As AI models get better at finding vulnerabilities at scale, expect more demonstrations like OpenAI's, and expect regulators to ask harder questions about when organisations knew and what they did. The evidence of that will matter as much as the response itself.