AI Is Exposing Your Data: An AI Security Problem You Can't See

Summary

The video explains how AI can increase data-exposure risks because sensitive information may pass through many connected components, often beyond the visibility of traditional security tools. It cites a study finding that 31% of organizations had experienced an AI-related data-privacy violation.

Where exposure can happen

• Shadow AI: Employees may deploy unauthorized AI tools without the security controls required by organizational policy.
• Public chatbots: Employees might submit sensitive files or information. Depending on the service and its settings, that information could be retained or used in ways the organization does not control.
• Complex AI workflows: Data may pass through training datasets, user prompts, RAG sources, policy and context instructions, models, agents, tools, databases, and other downstream systems. Agents may also launch additional agents.

Sensitive information can enter at multiple points, including training data, uploaded documents, prompts, internal policies, and connected tools. It can also be transformed as it moves, making it harder to detect with checks that look only for the original format.

What organizations need to monitor

The presenter separates visibility into two streams:

• AI workload stream: Inspect AI applications, RAG pipelines, and vector databases; track data transformations, sources, and destinations.
• Workforce stream: Monitor employee uploads and downloads, copy-and-paste activity, derived or “child” files, and how AI-generated or retrieved information is shared.

The goal is end-to-end data lineage: knowing where data came from, how it changed, where it went, whether its use was authorized and governed, and how to investigate potential exposure.

Discovery approaches and gaps

The video describes three complementary discovery perspectives:

1. Agentic-platform discovery: Shows which user prompted which model, which agent ran, and which MCP tool was called.
2. Endpoint DLP discovery: Identifies agents, extensions, MCP servers, files, memory, and other digital traces on devices.
3. Cloud or on-premises discovery: Finds data sources and can identify their classification, sensitivity, and ownership.

Each approach sees only part of the environment. The video argues for integrating them into a unified view with shared policies, monitoring, and proactive risk identification.

Recommended capabilities

A system for managing AI-related data exposure should provide:

• Continuous, AI-aware discovery and classification across platforms, including PII, health and financial information, and intellectual property.
• Lineage-based risk visibility that tracks transformations and propagation through AI systems.
• Context-aware investigation that considers the user, data sensitivity, destination, and other factors—reducing investigations from weeks to minutes.
• Compliance visibility for frameworks and regulations such as GDPR, the EU AI Act, SOC 2, ISO 27001, and HIPAA.

Content type and source

This is an explanatory overview of AI data-exposure risks and security capabilities. It does not review a specific product or provide a step-by-step tutorial.

Main speaker/source: Presenter Jeff Crume, from IBM Technology.