A control architecture for mature, regulated enterprises  already running Claude Code, OpenAI Codex, GitHub Copilot, and Cursor against production systems. 

Almost all of the AI security writing produced in 2026 is aimed at one question: is the code an AI wrote safe to ship? It is a fair question. We think it is the wrong one on which to build a control program. 

The evidence on code quality is not reassuring, and it deserves a paragraph before we set it aside. The foundational study of GitHub Copilot found roughly 40 percent of 1,689 generated programs contained a vulnerability. Four years and many model generations later, Veracode measured the security pass rate for AI-generated code holding flat at 56 percent across more than a hundred models.

Mature engineering organizations already know how to metabolize that class of risk. It is a code quality problem, and code quality problems route into static analysis, peer review, and release gates that have existed for two decades. What those controls need is retuning for volume, not reinvention. 

The harder problem lies in what a coding agent actually is. Treating it as a code generator understates it badly. It is a privileged actor operating inside your engineering estate, with credentials, network access, and the authority to execute. It reads material it did not author and cannot fully vet, acts on that material at machine speed, and in a single session touches more of the estate than most of the engineers supervising it. 

Securing the artifact is an application security problem, and one most organizations have already solved in substance. But securing the actor is a different matter: an identity, authorization, and detection problem that in most places we have looked has not been framed yet, let alone funded. 

Read our whitepaper for a comprehensive crash course on securing AI coding assistants.