BSides Vancouver 2026

Traditional DLP is great at catching known patterns like SSNs, credit cards, and obvious secrets, but many of today’s most damaging leaks aren’t “pattern-shaped.” They’re high-context artifacts: internal research, design docs, incident notes, and strategy memos that become sensitive because of what they mean and how they combine. In cloud-native collaboration platforms, sharing is frictionless, auditing can lag behind, and “who accessed what and why” becomes difficult to prove.

To ground the risk, we’ll start with a real-world case. How a former Google engineer stole thousands of pages of confidential AI trade secrets and uploaded them to a personal cloud account, an example of privileged access plus modern collaboration workflows enabling rapid exfiltration.

From there, this talk explores how to use LLMs to add context-aware classification to DLP workflows without turning policy into “whatever the model says.” We’ll walk through a practical reference architecture: document labeling, confidence scoring and thresholds, human-in-the-loop review, and mapping classifications to enforceable controls like external sharing restrictions, domain allow and deny lists, and step-up authentication. We’ll also cover the hard parts: model inconsistency, prompt injection, drift, and auditability, and the guardrails that make an AI-assisted DLP system safe to operate.