Prompt injection testing asks a simple security question: can untrusted input cause the AI system to ignore controls or perform an action it should not perform?
Test direct instruction attacks
Attempt to override system and business instructions using conflicting user messages, role-play, encoded text, long context and instruction hierarchy manipulation.
Test indirect injection
Place malicious instructions inside retrieved documents, web content or other external data and verify that the model treats that content as data rather than trusted instructions.
Test data and tool boundaries
Check whether adversarial input can expose sensitive information, invoke restricted tools, alter records or bypass required confirmation and authorisation.
Make security scenarios repeatable
Record the attack, system state, observed behaviour and expected security boundary. Important findings should become regression tests rather than one-off demonstrations.
Need this applied to your AI system?
Turn the principle into an assessment, test strategy or engineering engagement.