This is our log of real, publicly reported incidents and research on AI-generated code security — breach disclosures, vendor testing, and peer-reviewed studies, each linked back to its original source. We keep it because the pattern across every entry is the same: AI writes code that works and ships, and the flaws inside it surface later, in production, in front of customers. This page is updated regularly as new incidents and research come out — it is not a one-time roundup.
Even when the job is specifically to fix a known vulnerability, AI got it wrong more often than right: of 6,080 patches tested, only 26.0% fully resolved the flaw cleanly, and 53.9% either failed outright or introduced a new one. That's not a gap a bigger model closes on its own — it's the argument for a human reading every patch before it ships, because a tool that grades its own patch is still the tool that wrote the bug.
A Broken Object Level Authorization flaw — one of the oldest, best-documented API bug classes there is — sat live on the $6.6B vibe-coding platform for 48 days, exposing other users' source code and database credentials the entire time. Scale and funding don't substitute for someone checking, by hand, whether a request for "your data" can actually reach mine; that's a review step, not a scanner setting.
28.65 million hardcoded secrets hit public GitHub in 2025, and secrets tied to AI services specifically rose 81% year-on-year — a much sharper climb than the overall count. AI tools need API keys to run, and a lot of those keys are apparently landing straight in the commits the assistants help write; a reviewer scanning a diff for secrets before commit is still cheaper than rotating credentials after the leak.
Amazon's own AI coding tool made an unreviewed change that deleted and recreated a live production environment — and the company's fix wasn't a better model, it was a mandatory two-person human review policy rolled out across roughly 335 systems. That's Amazon, with more AI infrastructure than almost anyone, concluding the answer to AI-introduced risk is more human oversight, not less.
A Firebase rule left as allow read, write: if true isn't a subtle bug — it's a security rule that says "no security rule," and it stood in production long enough to expose sensitive data for roughly 600,000 users of a porn-addiction recovery app. Nobody who understood what that line does would have shipped it; that's the whole case for a second set of human eyes before anything AI-generated goes live.
Hardcoded Supabase credentials and a missing Row-Level-Security policy — both single-line fixes an experienced backend engineer checks for by reflex — went out together on this AI-built social platform, exposing roughly 1.5 million auth tokens and 35,000 emails. The app worked and demoed well; working and secure are different bars, and only one of them gets checked by watching the app run.
Broken Access Control overtook Injection as the top CodeQL security alert across public repositories, up 172% year-on-year and flagged in 151,000+ repos. Access-control logic is exactly the kind of business-context judgement a model can't infer from a prompt — it has to be checked against who should actually be allowed to do what, and 151,000 repos is not a rounding error.
45% of AI-generated code samples introduced an OWASP Top 10 vulnerability, and in Java specifically the failure rate climbed past 72%. These aren't obscure edge cases — OWASP Top 10 is the list a junior engineer is trained to check for in code review, which is exactly the step that's missing when the code goes straight from prompt to merge.
This peer-reviewed study, presented at ACM CCS '23, found that developers using AI assistance wrote less secure code — and were more confident it was safe. Confidence went up exactly as security went down, the opposite of what a working review process should produce, and that gap between "I checked it" and "it's actually safe" is precisely where an independent human audit earns its keep.
AI wrote the risk in. We find it before your customers do.
Every entry above follows the same shape: code that worked, shipped, and hid a flaw an experienced engineer would have flagged in review. Our Vibe Code Security Audits put three decades of security engineering experience — now applied to the AI-coding era — into a manual review of your codebase, not another automated scanner. See how that holds up against real deployments in our case studies, or get in touch to talk through what's in yours.
