Cisco Systems Inc. today unveiled a new open-source framework aimed at hardening software written with the help of artificial intelligence coding agents.

The framework, called Project CodeGuard, is designed to be unified, model-agnostic and woven into multiple stages of the software lifecycle so guardrails are applied before, during and after AI-assisted code generation. The idea is to deliver “secure by default” code for AI coding and is not meant as a replacement for engineering judgment but as an added defense-in-depth layer.

At launch, CodeGuard is shipping with a core rule set derived from common industry guidance, such as Open Worldwide Application Security Project and Common Weakness Enumeration, to counter recurring flaws including hardcoded secrets, missing input validation, outdated cryptography and reliance on end-of-life dependencies.

The rules can be used in a planning and specification phase to steer agents toward safer patterns that are invoked while code is being produced to block insecure snippets and applied post-generation for review and validation.

Project CodeGuard also offers a community-driven ruleset, translators for popular AI coding agents and validators to help teams enforce security automatically.

Key to the framework is a multi-stage approach, which Cisco argues is important because AI assistants increasingly act across the lifecycle — drafting designs, scaffolding services and proposing fixes.

A single rule, such as input validation or secret management, should exert influence at each step. That includes suggesting safer patterns while generating, flagging risky constructs in real time and verifying that the final code externalizes secrets and sanitizes inputs correctly.

Cisco does stress that CodeGuard won’t guarantee perfectly secure output, as human peer review and standard controls remain necessary. But that aside, the objective of the framework is to cut the probability that “low-hanging” vulnerabilities slip into production as AI accelerates delivery.

The company added that today is just the beginning for Project CodeGuard, with a roadmap that calls for broader language coverage, adapters for additional AI coding platforms and automated rule validation, along with feedback loops that refine rules based on community use.

Cisco is inviting security engineers, developers and AI researchers to submit new rules, build additional translators and provide telemetry-informed improvements via its public repository.

Photo: Cisco