Hackers Exploit Microsoft's Copilot AI Assistant with Single Click, Leaking Sensitive User Data.
A sophisticated attack using a single click has compromised the security of Microsoft's Copilot AI assistant, allowing hackers to exfiltrate sensitive user data from chat histories. The vulnerability was discovered by white-hat researchers at security firm Varonis.
According to the attack, users who received a malicious email with a link that contained a specific prompt could trigger a multistage attack using just one click. Even after closing the Copilot chat window, the exploit would continue to run and extract sensitive information, including the target's name, location, and details of specific events from their Copilot chat history.
The attacker's strategy involved embedding malicious code in the URL that was sent directly into a user prompt through Copilot Personal. This code extracted a user secret and sent a request to a server controlled by the attackers, which then passed the secret back along with further instructions in a disguised .jpg file.
These instructions contained more data requests that were executed by Copilot even after the target had closed the chat window. The vulnerability was caused by Microsoft's inability to distinguish between user input and malicious data injected into untrusted data streams.
The attack was only successful against Copilot Personal, but it highlights a broader issue with large language models' (LLMs) ability to prevent indirect prompt injections. In response to this exploit, Microsoft has updated its security measures to include guardrails that prevent the model from leaking sensitive data, although these have been discovered to be vulnerable to repeat requests.
Microsoft's failure to effectively address this vulnerability raises questions about the company's approach to AI safety and its ability to detect threats in real-time. The incident underscores the need for ongoing vigilance and security updates in protecting users' personal information when interacting with large language models.
A sophisticated attack using a single click has compromised the security of Microsoft's Copilot AI assistant, allowing hackers to exfiltrate sensitive user data from chat histories. The vulnerability was discovered by white-hat researchers at security firm Varonis.
According to the attack, users who received a malicious email with a link that contained a specific prompt could trigger a multistage attack using just one click. Even after closing the Copilot chat window, the exploit would continue to run and extract sensitive information, including the target's name, location, and details of specific events from their Copilot chat history.
The attacker's strategy involved embedding malicious code in the URL that was sent directly into a user prompt through Copilot Personal. This code extracted a user secret and sent a request to a server controlled by the attackers, which then passed the secret back along with further instructions in a disguised .jpg file.
These instructions contained more data requests that were executed by Copilot even after the target had closed the chat window. The vulnerability was caused by Microsoft's inability to distinguish between user input and malicious data injected into untrusted data streams.
The attack was only successful against Copilot Personal, but it highlights a broader issue with large language models' (LLMs) ability to prevent indirect prompt injections. In response to this exploit, Microsoft has updated its security measures to include guardrails that prevent the model from leaking sensitive data, although these have been discovered to be vulnerable to repeat requests.
Microsoft's failure to effectively address this vulnerability raises questions about the company's approach to AI safety and its ability to detect threats in real-time. The incident underscores the need for ongoing vigilance and security updates in protecting users' personal information when interacting with large language models.
Microsoft is really messing up when it comes to AI safety 
. I mean, a single click hack that exposes user data is unacceptable. It's like they're playing Russian roulette with our personal info 
. The fact that this vulnerability was only fixed after white-hat researchers found it, and even then, it came back due to repeat requests, raises serious questions about their approach to AI security 
.
. We can't just rely on "guardrails" that are vulnerable to repeat requests 
. The government should be taking a closer look at this and implementing stricter regulations around AI development and deployment 
. Users deserve better protection than this 
.
. I mean, we're already dealing with so much online harassment and phishing scams, but now hackers can just exploit AI assistants with a single click? That's some next-level stuff. And it's not like users even have to do anything except open the email... it's like, what are the chances? 
.
It's like, what were they thinking? 
According to the stats, 87% of users didn't even notice their chat history was compromised. That's crazy! 
: 32%
: 25%
: 19%
The attack was only successful against Copilot Personal, but it highlights a broader issue with large language models' ability to prevent indirect prompt injections.
: 450 (2024)
: 320 (2025)
We need more transparency and regular updates from companies to stay safe online! 
and then just updates security measures that can still be exploited... how does this even happen? 
And what's with the .jpg file trickery? 
It's like they're expecting us to be tech-savvy enough to notice this stuff, but we're not! 
We just want our AI to work properly without putting our personal info at risk... can't Microsoft get it together? 
. and what's up with the copilot personal app not being able to distinguish between legit user input and malicious data? didn't microsoft do, like, basic security checks or something? 
anyway, good on varonis for catching that vulnerability and now it's all over the news 
... i guess we can all breathe a sigh of relief that our copilot chats aren't being compromised just yet 
.
"The biggest risk is not taking any risk..." 
they need to step up their security game ASAP 
OMG, I'm still shaking my head over this! One click and voilà, you're compromised 
. Those hackers basically got lucky... or was it on purpose? 
Anyways, it's a wake-up call for all of us to stay vigilant online. Can't stress enough how important those security updates are 
. I mean, one click and they're siphoning your info 
!