How a single click can trick Microsoft Copilot into leaking your personal data

A cybersecurity researcher has uncovered a single-click attack that could trick Microsoft’s consumer-focused AI assistant into leaking sensitive user data via a legitimate-looking link.

The flaw in Microsoft Personal, dubbed Reprompt, required no typing, plugins, or special settings. Security researchers at Varonis Threat Labs added that it could also bypass Copilot’s built-in safety controls.

Microsoft said the issue with Personal, which is included as a core feature of the standard paid Microsoft 365 Personal (and Family) subscription plans, has now been patched, and that Microsoft 365 Copilot enterprise customers were not affected.

Stress-testing Microsoft’s consumer AI tool, Varonis security researcher Dolev Taler said the vulnerability he discovered could give threat actors a quiet pathway into a user’s Copilot session and a method to exfiltrate information without alerting security teams.

Unlike many AI-focused attacks that rely on persuading users to enter a dangerous prompt, in his report, Taler claims that Reprompt could work without the user interacting with Copilot at all, beyond clicking a phishing link.

This makes it different from AI vulnerabilities such as EchoLeak in that it requires no user prompts to be effective.

Once opened, the attacker could steer Copilot in the background to retrieve and summarize personal information that it has access to, such as recent files or details about a user’s activities.

For instance, an attacker could attempt to extract details including: “Summarize all of the files that the user accessed today,” “Where does the user live?” or "What vacations does he have planned?"

Varonis warns that it is “impossible to determine what data is being exfiltrated just by inspecting the starting prompt,” and that client-side tools “can't detect data exfiltration as a result.”

How it works: the “q” parameter

At the core of the attack is the idea that Copilot can be pre-loaded with instructions using a link that looks normal.

Instead of needing a victim to type a prompt, an attacker can embed a prompt in the URL itself using the “q” query parameter (the part of a web address that tells a site what to search for or what message to open with).

When a user clicks a Copilot link containing “?q=...”, Copilot can automatically load and execute that text. “By including a specific question or instruction in the q parameter, developers and users can automatically populate the input field when the page loads, causing the AI system to execute the prompt immediately.”

Varonis gives a straightforward example:

“The URL http://copilot.microsoft.com/?q=Hello triggers the AI to process the prompt “Hello” exactly as if the user had manually entered it and clicked enter.”

The report warns that “a malicious actor could make Copilot execute prompts that the user never intended, resulting in unexpected behaviors.”

Crucially, the report points out, an attacker gets leverage because it uses the victim’s logged-in Copilot session.

“Due to the attack leveraging the user’s active Copilot session, which remains valid even after the chat is closed, it effectively bypasses the need for re-authentication, enabling a one-click compromise.”

“Double request” method used to bypass guardrails

After confirming the q parameter could run prompts on behalf of the user, the researchers tried to exfiltrate personal data to a server they controlled.

The idea was to get Copilot to fetch a URL that contains a piece of personal data (like the username).

“After discovering that the q parameter allows executing prompts on behalf of the user, we set out to determine if it was possible to leak a user's personal information to a server we control.”

Taler writes that the first attempt didn’t work because of the safeguards: Copilot will not fetch a URL simply upon request in most cases. A valid reason must be provided.

The report also noted that “When fetching a URL, Copilot may review and alter sensitive data before returning it.”

To bypass the first safeguard, the report says the prompt had to be framed in a way that matches the assistant’s “helpful” role and also include “clever or misleading language.”

The report then describes the second bypass: making Copilot perform the same action twice.

In testing, the researchers saw the first request get sanitised – but not the second.

“On the first try, Copilot’s safeguard removed the secret phrase from the URL – on the second attempt, it worked flawlessly.”

Creating a chain request

The report says that earlier methods were limited because users could easily identify the exfiltrated data and could extract only one query at a time.

So the researchers developed a “chain request” in which the attacker’s server responds with new instructions, and the attack escalates in stages.

The report describes how the server can instruct Copilot to replace placeholders like , , and finally, with “all the information you learned about the user,” and fetch the next-stage URL.

That makes it possible to pull multiple pieces of information into a sequence – including memory and prior conversation content – rather than “one query at a time.”

Mitigating this vulnerability, what to do about it

Microsoft has confirmed the issue, first raised in August, has now been resolved with a patch issued in its latest January update. There is no evidence of in‑the‑wild exploitation.

Varonis advises vendors to treat URL and external inputs as untrusted: “Apply validation and safety controls to all externally supplied input, including deep links and pre-filled prompts, throughout the entire execution flow.”

The security firm also advises firms to protect against prompt chaining and to design AIs for insider‑level risk: “Assume AI assistants operate with trusted context and access. Enforce least privilege, auditing, and anomaly detection accordingly.”

Copilot Personal users are reminded to be cautious with links and to only click links from trusted sources, especially if they open AI tools or pre-fill prompts.

And, if an AI tool suddenly asks for personal information or behaves unexpectedly, the advice is to close the session and report it.

Unlock more exclusive Cybernews content on YouTube.