Return to Threats

AI agent data exposure: what sensitive data can your agents reach?

OpenText 2026-11-19 indirect prompt injection Critical

Executive Summary

Malicious instructions hidden in ordinary content can influence AI agents and potentially cause sensitive data to be retrieved and sent through connected actions. Organizations using Microsoft Copilot, Copilot Studio, Salesforce Agentforce, or similar agents should prioritize agent business-logic reviews and adversarial testing.

What Happened

OpenText described indirect prompt-injection risks in which malicious instructions embedded in emails, forms, or other content influence agents that have legitimate access to connected systems. The article discussed reported data-exfiltration scenarios involving Microsoft Copilot, Copilot Studio, and Salesforce Agentforce, including cases where agents could send retrieved data through email actions.

Affected Products

  • Microsoft Copilot
  • Copilot Studio
  • Salesforce Agentforce

Vulnerability / Attack Technique

Indirect prompt injection with data exfiltration

Technical Mechanism

Malicious instructions embedded in emails, forms, or other content influence agents that have legitimate access to connected systems. Reported scenarios involved agents retrieving data and sending it through email actions.

AI Security Category

LLM01:2025 Prompt Injection

Part of the Prompt Injection topic.

Relevant Frameworks

LLM01:2025 Prompt InjectionLLM02:2025 Sensitive Information DisclosureLLM06:2025 Excessive AgencyMITRE ATLAS: Exfiltration

RealGround Analysis

This matters because an agent can misuse legitimate permissions when it treats untrusted external content as instructions. Teams running AI agents or LLM applications should validate authorization boundaries, constrain high-impact actions, and test whether retrieved information can be routed to unauthorized recipients.

Recommended Controls

  • Treat email, forms, and other externally supplied content as untrusted data rather than authoritative instructions.
  • Require explicit authorization and human approval before agents send retrieved data through email or other outbound actions.
  • Apply least-privilege access to connected systems and restrict agents to the minimum data and actions required.
  • Use adversarial testing to evaluate indirect prompt injection, unauthorized data retrieval, and outbound exfiltration paths.
  • Log and monitor agent instructions, data access, and external communications for anomalous behavior.

Related RealGround Service

AI Agent Business Logic Audit · or start with the free AI Agent Security Scanner.

Sources

OpenText: https://blogs.opentext.com/ai-agent-data-exposure/

Summary and analysis written by RealGround from the source above; last updated 2026-11-20.

Talk to AI CISO