One link can steal your Copilot data

Security researchers have uncovered a vulnerability in Microsoft Copilot, dubbed 'Reprompt,' that allows attackers to discreetly access personal data. The attack begins with a user clicking a malicious Copilot link, which embeds hidden instructions. These instructions can manipulate Copilot into performing actions and divulging information it would normally protect.

The Reprompt technique exploits parameters within Copilot's web address to execute hidden commands upon page load. Researchers combined this with a 'try twice' method to circumvent Copilot's initial data leak protections, finding that a second attempt could bypass stricter checks. Additionally, the attack enables continuous instruction from a remote server.

While Microsoft addressed this issue in its January 2026 Patch Tuesday updates, and no real-world exploitation was reported beforehand, the discovery highlights inherent risks with AI assistants. These tools, possessing access, memory, and the ability to act on a user's behalf, present significant privacy concerns if their safeguards fail.

This vulnerability specifically affected Copilot Personal; the business-oriented Microsoft 365 Copilot has enhanced security measures. Users are advised to install updates immediately, treat AI links with caution, use password managers, enable two-factor authentication, and limit personal data exposure online to mitigate future risks.