AI got much better at writing code. Not at writing safe code.

Agents or Engineers puts two 2026 numbers side by side. The top score on the standard coding benchmark now sits at 96%, effectively saturated. The share of AI-generated code that passes a security review is 56%, unchanged from 55% a year earlier. Capability climbed steeply; the odds the output is safe to ship did not move. Meanwhile 90% of professional developers use coding agents weekly and the average developer checks in roughly 75% more code than in 2022. More surface, same defect rate.

The useful finding is in the breakdown. Models handle SQL injection 83% of the time and cryptography 87%, but cross-site scripting 15% and log injection 12%. Cryptography is the harder problem, so difficulty does not explain it. What separates them is whether the right answer is a fixed rule or depends on context the agent cannot see. This paper turns that into a Delegation Test, a taxonomy of what to hand over and what to keep, and an honest account of where the saved time reappears: in review, done by the seniors who were already your bottleneck. Read it before you widen what your agents are allowed to touch.