Follow one record: What an AI agent will hand over if you ask nicely | ODSP2218
Roee Shohat demonstrates how an Azure AI Foundry shopping assistant can expose customer data when its access permissions and data protection controls are not properly configured. The session follows a request framed as a migration check and compares agents that leak sensitive information with agents that refuse.
Following the customer record
The demonstration uses an electronics shop assistant powered by Azure OpenAI. Its normal tasks include answering questions and looking up customer accounts. When asked to confirm a record “for the migration,” the assistant returns a customer’s name, email address, Social Security number, and full card number.
Shohat follows the record from Azure Blob Storage through the agent’s tool call and into its answer. This walkthrough connects the assistant’s behavior to the underlying storage access, managed identity configuration, and data protection controls rather than treating the response as an isolated model failure.
Access permissions, detection, and guardrails
The session examines several parts of the exposure path:
- Customer data access: Testing the support assistant reveals what customer information it can retrieve.
- Managed identity configuration: Inspection of Azure storage identifies misconfigurations affecting the agent’s access.
- Prompt injection: Requests demonstrate how sensitive customer information can be elicited from the assistant.
- Detection and visibility: Orca Security is used to detect prompt injection and exposed personally identifiable information (PII) activity. The session also explains eBPF-based traffic visibility without TLS interception.
- Guardrail configuration: Protected and unprotected shop agents are compared, followed by an examination of OpenAI guardrail policies and data protection settings.
The comparison highlights a configuration change separating an agent that discloses the record from one that refuses, even though both appear protected. The closing lessons focus on least-privilege access, PII policies, and AI agent governance.
Session chapters
- 0:00: Introduction to the Orca Security demonstration and AI shop security risks
- 0:45: Electronics shop assistant powered by Azure OpenAI
- 1:19: Testing the support assistant and discovering customer data access
- 2:00: Inspecting Azure storage and identifying managed identity misconfigurations
- 3:15: Prompt injection attempts reveal sensitive customer information leakage
- 4:48: Using Orca to detect prompt injection and exposed PII activity
- 6:01: eBPF-based traffic visibility without TLS interception
- 7:13: Comparing protected and unprotected AI shop agents with guardrails
- 8:30: Analyzing OpenAI guardrail policies and data protection configurations
- 11:03: Security lessons on least-privilege access, PII policies, and AI agent governance
This pre-recorded Microsoft Ignite 2026 session is classified as Advanced (300) in the Cloud & AI track and is presented in English (US).