AI guardrails play an important role in helping prevent unsafe, non-compliant or unsupported outputs. But as AI models and the ways people interact with them continue to evolve, guardrails need to evolve too. Static test sets and hand-written adversarial prompts cannot anticipate every possible interaction.
This interactive demonstration shows how CommBank AI Labs uses adversarial testing to proactively identify potential weaknesses in AI guardrails so they can be understood and strengthened. Attendees can test a privacy guardrail in a controlled environment and see how different prompts challenge the model’s ability to identify and protect private information.
The experience is based on CommBank AI Labs’ research into self-improving adversaries. Alongside human testing, the demonstration introduces an automated adversary that learns to generate increasingly challenging tests, showing how reinforcement learning can extend red teaming beyond fixed prompt libraries and help uncover behaviours that conventional evaluation may miss.
The demonstration also explores the other side of AI safety: a guardrail can be highly restrictive while also blocking legitimate requests. By examining both potential policy violations and over-refusal, the experience shows why effective evaluation needs to consider safety, usefulness and customer impact together.
Attendees will leave with a practical understanding of how adversarial testing, independent evaluation and human review can help organisations continuously test and strengthen AI guardrails as models and threats evolve.
The demonstration will use synthetic data only. No real customer or personal information will be used as part of the demonstration.