Indirect Prompt Injection Hits 1.2M Web Pages, Ranked Top AI Vulnerability in 2026

Indirect Prompt Injection (IPI), ranked the number-one LLM vulnerability by OWASP, has been detected on over 1.2 million public web pages as of 2026, with a 32% growth rate according to Forcepoint's Global AI Threat Landscape Report. Unlike traditional cyberattacks, IPI requires no system breach — malicious instructions are hidden inside ordinary web content that AI agents read and unknowingly execute. A real-world example, EchoLeak (CVE-2025-32711), saw attackers exploit Microsoft 365 Copilot through a crafted email, causing it to automatically forward users' calendar data and contacts to the attacker, affecting over 100,000 users. The threat is compounded by Model Context Protocol (MCP), an AI integration layer whose architectural design allows malicious tool descriptions to inject override instructions directly into an agent's context window. Because affected agents often have access to tools like email, databases, and APIs, a single successful injection can enable data exfiltration, order manipulation, or phishing attacks launched from a trusted company domain.