AI writes more defects than review can catch

Veracode tested over a hundred models across eighty coding tasks and found security problems in about 45 percent of the output. A comparison of 320 pull requests written with AI against 150 written by people counted 1.5 times as many of them, against 1.7 times as many issues overall.

Across model generations that measurably improved on coding benchmarks, the security pass rate stayed flat. Whatever is driving those functional gains isn't producing safer output.

When findings go up, the reflex is more review capacity, because that's the control the volume appears to have outgrown. Well structured review can catch most of what ordinary use would have surfaced. It misses the defects that nothing surfaces until somebody tries to break in.

Normal development doesn't produce feedback about being attacked.

On the enterprise systems I worked on in insurance and healthcare, a sprint gave you a window to implement, fix and validate an item. When one took longer than estimated, the hours went into getting the logic right, then defensive programming and null checks, then testing against varying levels of correct data. Negative testing came last, and time decided how many of those cases ran. Agent assisted development may have changed that.

Negative testing also works from the specification, so it covers format, maybe length, maybe a header check. Those are the testing inputs somebody thought to write down.

A specification doesn't tell you to embed executable code inside a format you already accept. People keep finding model files on public hubs with a payload inside them. Each one is valid by the spec, and the loader that reads it runs whatever the stored bytes say to run. More of the same type of review isn't the answer.

What I'd fund is an adversarial capability inside the deployment pipeline, rather than an engagement booked twice a year. That kind of engagement gives you a snapshot of your system that changes with the next deployment. Build it AI assisted, because enumerating the published record of vulnerabilities against your specific packages and versions is the part that got cheap. Researchers ran that at about three and a half dollars a run.