Securing AI with strict boundaries

Once your agents can reach real systems and act on live data, the question stops being “can it do this?” and starts being “should it, and who’s accountable if it doesn’t stop itself?” This session digs into the consequences of not properly implementing the right boundaries and guardrails for AI agents. We start with the Moffatt v. Air Canada case, a chatbot that invented a bereavement fare policy that didn’t exist, as a leading example. We explore the full landscape of agent risk: hallucinated policy, PII leakage, prompt injection, jailbreak attempts, and the more mundane but still damaging problem of an agent wandering into an off-topic, embarrassing tangent.

From there, we get concrete about the guardrail mechanisms that actually hold these boundaries: topical guardrails that draw a hard line around specific subjects, ethical moderation layered on top of a model provider’s defaults, jailbreak detection, PII redaction with support for custom entities, and groundedness, forcing an agent to answer only from the context it’s been given.

We’ll also unpack why compliance certifications like SOC 2 or HIPAA, while real and meaningful, answer a different question than “does this agent follow the specific policy my legal team wrote,” and why that policy needs to be authored and enforced inside your own environment, not trusted to a vendor’s infrastructure.