Is Vibe Coding Actually Dangerous? What Security Researchers Are Saying in 2026

Security researchers spent most of 2025 saying "be careful." By 2026, some of them started reaching for a different comparison.

David Mytton, founder of developer security company Arcjet, predicted 2026 would bring "big explosions" as more unreviewed vibe-coded applications hit production. Simon Willison, a well-known figure in the developer tooling world, went further, comparing the moment to the Challenger disaster — not because the analogy is exact, but because of what it points at: normalized risk that nobody addresses until something breaks catastrophically. Willison specifically flagged how casually developers grant coding agents broad system access: "my computer doesn't get wiped [so] I'm like, 'oh, it's fine'" — until, statistically, it isn't.

The numbers back up the concern more than the rhetoric might suggest. Georgia Tech's Vibe Security Radar tracked new CVE entries directly caused by AI-generated code jump from 6 in January 2026 to 35 in March — roughly a sixfold increase in two months. Separately, when Veracode tested whether AI models would choose a secure or insecure coding pattern when both were available, models picked the insecure option 45% of the time — close to a coin flip.

There's also a structural problem, not just a volume one. A May 2026 industry report on agentic coding security noted that standard code-scanning tools (SAST, SCA) don't have a detection category for a growing class of vulnerabilities living in agent configuration files and instruction sets rather than the code itself — meaning some of the newest attack surface isn't covered by the security tooling most teams already have in place.

What this means if you're not shipping enterprise software

Most individual developers aren't running a platform with 18,000 users. But the underlying skill gap is the same one at any scale: can you actually evaluate whether AI-generated code is safe, or are you trusting that it probably is because it runs? The industry-level incidents are a magnified version of exactly what the test measures at an individual level — verification habits that haven't kept pace with how much code is being generated.

The response from the industry, per multiple 2026 reports, is heading toward more review requirements and tighter agent permissions by default — not less AI usage, just less blind trust in what it produces.

See where your own verification habits land →

Sources: The New Stack reporting on Mytton and Willison's comments (January 2026); Georgia Tech's Vibe Security Radar; Veracode's model security testing; industry security reporting, spring 2026.