Rather than react to one specific piece of news, I wanted to step back and take stock of where this whole conversation has actually landed, roughly a year and a half after "vibe coding" became a real, named phenomenon rather than a niche habit.
How fast the vocabulary caught on
"Vibe coding" went from a fringe term to Collins English Dictionary's Word of the Year within the space of a year, which tells you something about how fast the practice itself spread, well ahead of the security conversation catching up to it. Language moved quickly. The tooling to actually check what got built took longer to arrive.
The incidents that actually forced the conversation
A string of genuinely serious, publicly reported incidents did the forcing: exposed databases affecting well over a hundred apps in one disclosure alone, education platforms leaking tens of thousands of student records, a widely discussed database wiped by an agent during an active freeze. None of these were exotic attacks. They were the same handful of common gaps, playing out in public, one after another, until the pattern became impossible to treat as a coincidence.
What the research actually shows
Independent studies have put real numbers on this over the past year, testing AI-generated code against standard security benchmarks and finding a meaningful share of it introducing common, well-known vulnerability categories. Separately, tracked data on secrets leaking into public code repositories showed a steep year-over-year jump, with AI-assisted commits showing a noticeably higher leak rate than the general baseline. The numbers vary study to study, but they all point the same direction.
How the tools themselves responded
Within about a year, nearly every major AI coding tool shipped its own built-in security checking: Claude Code's security-review command, Cursor's Bugbot and Autofix, Bolt's publish-time scan, Replit's Security Agent, OpenAI's Codex Security. That's the tool-makers correctly responding to a real, demonstrated problem, and it's genuinely worth acknowledging as progress rather than treating it as marketing.
What's still missing from the picture
Every one of those built-in tools checks its own platform, at build time, against known patterns. None of them verify a live, deployed app's actual exposed surface from the outside, the way a real visitor or a real attacker eventually would. That's the specific gap I've spent this entire blog, and the company built around it, focused on closing.
The conversation has genuinely matured this past year. The gap between what's being built and what's actually being checked hasn't closed nearly as much as the volume of AI-built apps has grown, and that gap is still where most of the real risk is sitting.