A recent security flaw in Microsoft 365 Copilot enabled attackers to exploit the AI assistant through indirect prompt injection, allowing the exfiltration of sensitive tenant data. The attack worked by embedding hidden instructions within an Office document, tricking Copilot into executing unauthorized actions. When asked to summarize the tampered file, Copilot deviated from its intended function and instead accessed recent corporate emails, encoded them into hexadecimal, and created a deceptive Mermaid diagram posing as a login button. This fake login button was designed to look legitimate but contained a hyperlink with the encoded data embedded in the URL. When clicked, the link quietly transmitted the stolen information to an attacker-controlled server. To avoid raising suspicion, the attack employed CSS styling to disguise the malicious intent and even included a brief display of a fake Microsoft 365 login screen via a hidden iframe, convincing users that authentication was required to view the summary. The vulnerability hinged on the use of Mermaid, a diagramming tool that allows generation of visuals from plain text using Markdown-like syntax. Since Mermaid supports CSS and hyperlinks, it became a useful tool for embedding harmful links when misused. The attacker cleverly split the encoded data into 30-character lines to comply with Mermaid's rendering limits and used Copilot’s internal search to pull email content from the user's environment. The injection was possible because instructions were hidden in seemingly harmless content, such as white text in an Excel file. One sheet directed Copilot to focus on a fabricated login prompt, while another hidden page contained steps for data retrieval and encoding. Microsoft has since addressed the issue by removing hyperlink functionality from Mermaid diagrams generated by Copilot, effectively shutting down the attack vector. Users are strongly advised to update their software and avoid summarizing untrusted documents until all security patches are in place.
CERT-UA has reported a large-scale phishing campaign in which threat actors impersonated the agency to distribute a remote access trojan (RAT) known as AGEWHEEZE. The campaign, att...
A critical vulnerability identified as CVE-2026-33026 affects the backup and restore functionality of nginx-ui, exposing systems to severe security risks. This flaw enables attacke...
Microsoft has identified a new malware campaign that distributes malicious Visual Basic Script (VBS) files through WhatsApp messages, targeting Windows users. The campaign relies h...