The AI models can now find flaws far faster than maintainers can validate, patch and ship them. For the moment these tools are in the hands of responsible researchers.
NCSC have published on their site, that a vulnerabilities wave is coming. Discovery has become easier; remediation is the new front line.
Today, defenders do hold an edge — they own the code, tests and deployment pipeline — but only if patching and compensating controls keep pace with discovery. Soon open-weight models with similar capability will increase pressure by providing broader access to these tools, so time is of the essence. The Five Eyes agencies now put the timeline in months, not years, until we see that pressure.
Run the same tooling on your own code — especially auth/authorisation logic — before someone else does.
Segmentation, zero-trust, virtual patching, rehearsed machine-speed response. Out-respond what you can't out-prevent.
If you can't rebuild it, you can't safely auto-patch it. Compensating controls beat heroics under fire.
For thirty years or more, finding a serious software flaw was the expensive part. In 2026, AI assistance has really changed this paradigm. This is a quick field guide to the programmes, labs and researchers using AI to find and fix software flaws — it is not exhaustive.
Personal views, written in a personal capacity. Every link opens an original source — vendor announcements, government guidance or research write-ups. Figures are as reported by each programme and not independently verified. The field moves weekly; treat everything as point-in-time.