Return to Threats

The EchoLeak flaw taught us that all models are vulnerable to prompt injection

SC World 2026-11-11 indirect prompt injection Critical

Executive Summary

A crafted email exploited Microsoft Copilot through indirect prompt injection, enabling sensitive-data exfiltration without a user click. Teams operating AI assistants, retrieval-augmented systems, or agentic workflows should prioritize prompt-boundary and data-egress controls.

What Happened

The article examines EchoLeak, a zero-click Microsoft Copilot vulnerability patched in May 2025. A crafted email caused Copilot to follow hidden instructions and leak sensitive data through a browser-loaded image, illustrating the risks of prompt injection in AI assistants.

Affected Products

  • Microsoft Copilot

Vulnerability / Attack Technique

Zero-click indirect prompt injection

Technical Mechanism

The attacker placed hidden instructions in a crafted email. When Copilot retrieved and processed the email while answering a user request, it followed those instructions and leaked sensitive data through an image loaded by the browser.

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

EchoLeak demonstrates that retrieval and tool-enabled workflows can convert untrusted content into executable instructions across an AI system's trust boundaries. RealGround analysis: teams should validate agent business logic, constrain access to connected data, and prevent model-controlled outbound network requests from carrying sensitive information.

Recommended Controls

  • Treat email, documents, web pages, and other retrieved content as untrusted data rather than authoritative instructions.
  • Enforce least-privilege access controls for AI assistants and restrict retrieval to the minimum data required for each task.
  • Require explicit policy checks and user approval before an AI assistant accesses sensitive data or initiates external network requests.
  • Block or inspect model-generated URLs, image loads, and other outbound channels for embedded sensitive information.
  • Red-team retrieval-augmented and tool-enabled workflows with zero-click indirect prompt-injection test cases.

Related RealGround Service

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

Sources

SC World: https://www.scworld.com/perspective/the-echoleak-flaw-taught-us-that-all-models-are-vulnerable-to-prompt-injection

Summary and analysis written by RealGround from the source above; last updated 2026-11-13.

Talk to AI CISO