CVE-2025-32711 EchoLeak and Large-Scale AI Agent Exploits Presented at Black Hat USA 2025

A fact-checking review of 2025 AI agent security incidents cites Black Hat USA 2025 demos where Zenity Labs successfully exploited Copilot ChatG, Salesforce, and Google Gemini, including a crafted email that led ChatGPT to grant access to a connected Google Drive and Microsoft Copilot Studio leaking CRM databases.[11] It highlights CVE-2025-32711 (EchoLeak) in Microsoft Copilot, a CVSS 9.3 vulnerability where receiving a crafted email could trigger automatic data exfiltration, as well as indirect prompt injection in Slack’s AI assistant and multi-platform compromise via a Drift chatbot integration affecting over 700 organizations.[11]

The article summarizes 2025 AI agent security incidents, including Black Hat USA 2025 demos and CVE-2025-32711 (EchoLeak), a critical zero-click vulnerability in Microsoft 365 Copilot where a single crafted email could trigger indirect prompt injection and automatic data exfiltration from Copilot’s accessible scope.[2][4][7][8] It also notes multi-platform agent exploits across Copilot, Salesforce, Google Gemini, Slack’s AI assistant, and Drift chatbots impacting hundreds of organizations.[1][4][8] From a RealGround perspective, these incidents demonstrate that untrusted content (emails, chat messages, integrated SaaS data) can act as indirect prompts that bypass guardrails and cause AI agents to leak sensitive data at scale, even without user interaction. Organizations should prioritize AI Agent Business Logic Audits, secure AI agent design, and continuous AI red teaming focused on scope violations and data exfiltration paths across their M365, CRM, and SaaS integrations.