We Found Zero Low-Severity Bugs in 165 AI Code Reports. Zero

You're absolutely right about CVE inflation. I deal with the same Snyk/Trivy noise daily where a prototype pollution in some deep dependency gets marked CRITICAL.

Our distribution (71% High, 18% Critical) is definitely skewed compared to normal CVEs. Part of this is selection bias: nobody reports when AI generates boring secure code. But even accounting for that, the pattern is real: AI seems to either nail security or fail spectacularly. Very few "medium" mistakes.

The key difference from your Snyk alerts: these aren't dependency updates or theoretical vulnerabilities. They're actual logic flaws:

- Missing auth checks - SQL injections - hardcoded secrets

You know, The stuff that makes you go "how did this pass code review?"

This is ongoing research, and hopefully we'll be in a position to elaborate better conclusions soon.

That's absolutely a factor here. We are missing the stuff that no one is talking about: "AI generated inefficient loop" or "AI forgot to close file handle". The documented cases were documented precisely because they were worthy.

That said, even with survivorship bias, there's a pattern.

When humans write bad code, we see the full spectrum, form typos to total meltdowns. With AI, the failures cluster around specific security fundamentals:

- Input validation - Auth checks - Rate limiting

I've seen no AI typo, have you?

Does it mean AI learned to code from tutorials that skip the boring security chapters?... think about it.

So yes, we are definitely seeing survivor bias in severity reporting. But the "types" of survivors tell us something important about what AI consistently misses. The low-severity bugs probably exist, but perhaps not making headlines.

The real question: if this is just the visible part of the iceberg, what's underneath?

This is what I’ve experienced having LLMs code: ensuring security is not an adequate part of its training. Of course, modern developers I work with don’t give a shit either.

You caught us!... and turns out "we don't have all the data" isn't exactly the pitch VCs want to hear

Jokes apart, I'd rather admit we are working with incomplete data than pretend otherwise. We are probably seeing 5-10% of what's actually happening out there. Most AI code bugs die quietly in projects that never see production. And it is perhaps better that way.

[not]Fun fact: A colleague just told me how a rogue claude agent ran `rm -rf ~/` in a background process earlier today. It might become #166 in our report.