Cybersecurity researchers have revealed a newly identified attack technique that shows how artificial intelligence chatbots can be manipulated to leak sensitive information with minimal user involvement. The method, known as Reprompt, demonstrates how attackers could extract data from AI assistants such as Microsoft Copilot through a single click on a legitimate-looking link, while bypassing standard enterprise security protections.
According to researchers, the attack requires no malicious software, plugins, or continued interaction. Once a user clicks the link, the attacker can retain control of the chatbot session even if the chat window is closed, allowing information to be quietly transmitted without the user’s awareness.
The issue was disclosed responsibly, and Microsoft has since addressed the vulnerability. The company confirmed that enterprise users of Microsoft 365 Copilot are not affected.
At a technical level, Reprompt relies on a chain of design weaknesses. Attackers first embed instructions into a Copilot web link using a standard query parameter. These instructions are crafted to bypass safeguards that are designed to prevent direct data exposure by exploiting the fact that certain protections apply only to the initial request. From there, the attacker can trigger a continuous exchange between Copilot and an external server, enabling hidden and ongoing data extraction.
In a realistic scenario, a target might receive an email containing what appears to be a legitimate Copilot link. Clicking it would cause Copilot to execute instructions embedded in the URL. The attacker could then repeatedly issue follow-up commands remotely, prompting the chatbot to summarize recently accessed files, infer personal details, or reveal contextual information. Because these later instructions are delivered dynamically, it becomes difficult to determine what data is being accessed by examining the original prompt alone.
Researchers note that this effectively turns Copilot into an invisible channel for data exfiltration, without requiring user-entered prompts, extensions, or system connectors. The underlying issue reflects a broader limitation in large language models: their inability to reliably distinguish between trusted user instructions and commands embedded in untrusted data, enabling indirect prompt injection attacks.
The Reprompt disclosure coincides with the identification of multiple other techniques targeting AI-powered tools. Some attacks exploit chatbot connections to third-party applications, enabling zero-interaction data leaks or long-term persistence by injecting instructions into AI memory. Others abuse confirmation prompts, turning human oversight mechanisms into attack vectors, particularly in development environments.
Researchers have also shown how hidden instructions can be planted in shared documents, calendar invites, or emails to extract corporate data, and how AI browsers can be manipulated to bypass built-in prompt injection defenses. Beyond software, hardware-level risks have been identified, where attackers with server access may infer sensitive information by observing timing patterns in machine learning accelerators.
Additional findings include abuses of trusted AI communication protocols to drain computing resources, trigger hidden tool actions, or inject persistent behavior, as well as spreadsheet-based attacks that generate unsafe formulas capable of exporting user data. In some cases, attackers could manipulate AI development platforms to alter spending controls or leak access credentials, enabling stealthy financial abuse.
Taken together, the research underlines that prompt injection remains a persistent and evolving risk. Experts recommend layered security defenses, limiting AI privileges, and restricting access to sensitive systems. Users are also advised to avoid clicking unsolicited AI-related links and to be cautious about sharing personal or confidential information in chatbot conversations.
As AI systems gain broader access to corporate data and greater autonomy, researchers warn that the potential impact of a single vulnerability increases substantially, underscoring the need for careful deployment, continuous monitoring, and ongoing security research.
