A tester’s guide to AI guardrails

AI systems introduce a new kind of risk. 

They don’t just fail; they produce plausible, unsafe, or misleading outputs while appearing correct. 

Guardrails are used to control these behaviours. But in most systems, they are: 

•  vaguely defined 
•  poorly tested 
•  incorrectly implemented 

In this session, Rahul Parwal introduces a structured approach to learning AI guardrails through an interactive, scenario-based format.

Participants will work through short exercises that reflect real testing challenges:

• Identifying types of guardrail failures 
• Determining when a guardrail should trigger 
• Recognising common attack patterns 
• Improving weak system prompt rules 
• Spotting implementation-level issues

By the end, participants will have a practical framework to test AI systems more systematically. 

Learning outcomes

• Understand the  various categories of AI guardrails 
• Practice a set of practical techniques to test AI guardrails 
• A clearer understanding of where set guardrails fail in real systems