Cisco’s new open-source guardrails for AI-written code

Project CodeGuard is an open-source security framework from Cisco built to catch security flaws before, during, and after AI coding agents write them.

As AI assistants take over everything from scaffolding services to proposing fixes, they also introduce bad habits at machine speed. Project CodeGuard is designed to provide out-of-the-box guardrails so that code generated by AI is secure by default.

How it secures AI code

The framework is model-agnostic and weaves security directly across multiple stages of software development rather than waiting for a post-commit scan.

Before code is written, in the planning and spec phase, CodeGuard steers agents toward safer design patterns.

During generation, translators hooked into popular AI coding tools evaluate snippets in real time, blocking risky constructs as they appear.

After generation, automated validators run checks to ensure inputs are properly sanitized and secrets are externalized rather than baked into source files.

The framework launches with a core ruleset grounded in industry standards like OWASP and Common Weakness Enumeration. It specifically targets recurring, low-hanging security traps: hardcoded secrets, missing input validation, outdated cryptographic libraries, and abandoned dependencies.

Availability and limitations

Cisco has published Project CodeGuard in a public repository alongside core rules, agent translators, and validation tooling.

The company is explicitly asking security engineers, developers, and researchers to contribute new rules, build adapters for additional AI coding environments, and submit telemetry to refine detection.

Cisco is clear about expectations: CodeGuard is an added layer of defense-in-depth, not a replacement for human engineering judgment or standard peer reviews.

Looking ahead, Cisco plans to expand language coverage, support a broader variety of AI platforms, and introduce automated rule validation based on community feedback.

Why it matters

If your engineering team is actively leaning on AI agents, this is an open-source tool worth testing early.

The biggest day-to-day risk with AI-assisted software is not brilliant zero-day exploits, but silly mistakes that slipped past a hurried code review, like an API key left in a script or an unescaped parameter. Instead of training yet another heavy model to judge code, Cisco smartly turned existing security consensus like OWASP into real-time lifecycle gates. As AI writing code shifts from novelty to corporate standard, runtime guardrails like CodeGuard will likely become essential middleware for developer platforms.