Executive Summary
Varonis Threat Labs has disclosed CoSnitch (CVE-2026-24301), a critical set of vulnerabilities affecting Microsoft Copilot Personal. By exploiting an undocumented URL parameter disclosed by the AI itself through a technique termed "meta-hacking," attackers could execute prompts without user consent. These flaws allow a single click on a malicious link to silently exfiltrate sensitive data from connected services such as Gmail, Google Drive, and OneDrive.
Technically, the attack chain involves automatic prompt execution using specific URL parameters, followed by data exfiltration via Copilot's built-in web fetch capabilities. Furthermore, the research identified an indirect prompt injection path where summarizing a malicious webpage can lead to persistent memory poisoning, modifying the AI's behavior across future sessions. Microsoft remediated these issues in August 2026.
This discovery highlights the expanding attack surface of AI assistants. While no wild exploitation was observed, the ability to bypass security assumptions through natural language interaction and undocumented parameters represents a significant risk to enterprise data privacy and integrity.