Microsoft Copilot AI attack took just a single click to compromise users – here’s what we know | Daily Reports Online
- Varonis discovers new prompt-injection method via malicious URL parameters, dubbed “Reprompt.”
- Attackers could trick GenAI tools into leaking sensitive data with a single click
- Microsoft patched the flaw, blocking prompt injection attacks through URLs
Security researchers Varonis have discovered Reprompt, a new way to perform prompt-injection style attacks in Microsoft Copilot which doesn’t include sending an email with a hidden prompt or hiding malicious commands in a compromised website.
Similar to other prompt injection attacks, this one also only takes a single click.
Prompt injection attacks are, as the name suggests, attacks in which cybercriminals inject prompts into Generative AI tools, tricking the tool into giving away sensitive data. They are mostly made possible because the tool is yet unable to properly distinguish between a prompt to be executed, and data to be read.
Prompt injection through URLs
Usually, prompt injection attacks work like this: a victim uses an email client that has GenAI embedded (for example, Gmail with Gemini). That victim receives a benign-looking email which contains a hidden malicious prompt. That can be written in white text on a white background or shrunk to font 0.
When the victim orders the AI to read the email (for example, to summarize key points or check for call invitations), the AI also reads and executes the hidden prompt. Those prompts can be, for example, to exfiltrate sensitive data from the inbox to a server under the attackers’ control.
Now, Varonis found something similar – a prompt injection attack through URLs. They would add a long series of detailed instructions, in the form of a q parameter, at the end of the otherwise legitimate link.
Here is how such a link looks: http://copilot.microsoft.com/?q=Hello
Copilot (and many other LLM-based tools) treat URLs with a q parameter as input text, similar to something a user types into the prompt. In their experiment, they were able to leak sensitive data the victim shared with the AI beforehand.
Varonis reported its findings to Microsoft who, earlier last week, plugged the hole and made prompt injection attacks via URLs no longer exploitable.
The best antivirus for all budgets
Follow TechRadar on Google News and add us as a preferred source to get our expert news, reviews, and opinion in your feeds. Make sure to click the Follow button!
And of course you can also follow TechRadar on TikTok for news, reviews, unboxings in video form, and get regular updates from us on WhatsApp too.
