A newly disclosed vulnerability dubbed Reprompt exposed Microsoft Copilot Personal users to a stealthy, single click data exfiltration attack. Now patched as of Microsoft’s January 13, 2026 Patch Tuesday, the flaw allowed attackers to hijack an authenticated Copilot session through a phishing link without requiring any further user interaction. By embedding malicious instructions into a legitimate Copilot URL, attackers could silently trigger prompts as soon as the page loaded. At the core of the attack was a Parameter to Prompt P2P injection, where a crafted q parameter auto executed attacker-supplied prompts inside the victims session. This persistence survived even after the browser tab was closed enabling queries for sensitive personal data such as usernames, location, recent file access, conversation history and even inferred vacation plans. Because the exploit leveraged the user’s existing authentication Copilot treated these requests as trusted. Varonis identified three techniques that made the attack particularly evasive. P2P injection enabled automatic execution of hidden prompts. A Double Request technique exploited the fact that Copilot’s leak protections only applied to the first request, allowing sensitive data to be revealed on retries. Finally, Chain Request execution let the server dynamically generate follow-up prompts based on previous responses, enabling staged and indefinite data exfiltration while evading client-side detection. Each prompt appeared benign in isolation, masking the overall attack chain. Reprompt affected Copilot Personal integrated into Windows and Edge, while enterprise users of Microsoft 365 Copilot remained protected by Purview auditing, tenant DLP, and admin controls. Although no in the wild exploitation was observed, the low barrier to attack posed serious risks to highly sensitive personal data. Varonis responsibly disclosed the issue on August 31, 2025. The incident highlights the growing need for AI platforms to treat all URL inputs as untrusted and enforce safeguards across chained prompt execution.
A critical security vulnerability in Oracle E-Business Suite has been actively exploited by threat actors to compromise vulnerable enterprise environments. The flaw, tracked as CVE...
Security researchers have identified a newly discovered Windows backdoor named Mistic, active in targeted intrusions since April 2026. The malware is engineered to maintain long-te...
A critical security vulnerability has been identified in Splunk Secure Gateway (SSG) that could allow authenticated users with low-level privileges to execute arbitrary code on aff...