Working Surfacelink · 19 Nov 2026
03
Two identical stamping presses stand side by side, one fitted with a precise metal stencil plate turning out clean, even parts and the other with no plate turning out bent, ragged parts, while a figure in silhouette checks the plate with a ruler, the whole scene drawn with wide empty margins on every side.
LeadDev

AI coding rules matter more than your model

Michael Hill · 18 November 2026

How-to · artifacts · agents · Agent context files

Read the original
Takeaway

A benchmark of over 30 AI models found that curated security rule files cut risk in AI-written code by up to 70%, while vague instructions did little.

Summary

Michael Hill, managing editor of LeadDev, a publication for engineering leaders, reports the second AI Trust Index from Secure Code Warrior, a company that sells secure-coding training. It ran more than 30 AI models inside the setup developers use daily and built over 15,000 apps. The results suggest that the rules a team gives its AI models matter more for security than which AI model it buys.

Key points
  • The AI models produced between 0.24 and 3.66 vulnerabilities per 1,000 lines of code, a 15-fold gap.
  • Curated rule sets cut measured risk by up to 70%, while a line asking the AI model to write secure code made little difference.
  • A low-cost AI model without rules carried less risk than a flagship AI model costing more than 80 times as much per run.
  • In a LeadDev survey, 81% of 591 engineering leaders said AI tools have write access to their code, and 32% have no formal policy for AI agents.
Implication

An engineering team can treat its AI rule files as code: written for each language, reviewed, and retested when the AI model changes. The benchmark is Secure Code Warrior's own, and the company sells the kind of governance its results favour, as Hill notes.

Plays
Engineering

Write a file of security rules for the team's AI coding tools in each programming language the team uses, and test the rules again whenever the team changes AI model.

Engineering

Test candidate AI models on the team's own languages and frameworks before choosing a company-wide default.

Derived by Working Surface from the article; more in the Playbook. Source line: Treat your AI rules like code: write them for each language you use, review them, and retest whenever you change models.

Source issue

19 November 2026: as AI agents write most of a team's code, the human check moves from reading each pull request to owning the rules on which changes a person reads.