Guardrails: the controls you build yourself

The article argues that businesses adding AI features need to build their own security controls, because safeguards supplied by AI providers do not protect a company’s data, systems or customers. Security consultant Kit Barker says prompt injection can lead an AI system to follow malicious instructions in a user message or in material it later reads, such as a web page, email or CV.

Barker says developers should assume an AI agent may eventually be manipulated and limit what it can access or do. In his coding pipeline, agents can write code and open pull requests, but hold no production deployment credentials, cannot change merge checks or approve their own work; GitHub branch rules enforce the controls independently of the agents’ instructions.