Cybersecurity researchers have disclosed particulars of a brand new assault known as CometJacking concentrating on Perplexity’s agentic AI browser Comet by embedding malicious prompts inside a seemingly innocuous hyperlink to siphon delicate information, together with from related providers, like e-mail and calendar.
The sneaky immediate injection assault performs out within the type of a malicious hyperlink that, when clicked, triggers the sudden conduct unbeknownst to the victims.
“CometJacking reveals how a single, weaponized URL can quietly flip an AI browser from a trusted co-pilot to an insider menace,” Michelle Levy, Head of Safety Analysis at LayerX, mentioned in an announcement shared with The Hacker Information.
“This is not nearly stealing information; it is about hijacking the agent that already has the keys. Our analysis proves that trivial obfuscation can bypass information exfiltration checks and pull e-mail, calendar, and connector information off-box in a single click on. AI-native browsers want security-by-design for agent prompts and reminiscence entry, not simply web page content material.”
The assault, in a nutshell, hijacks the AI assistant embedded within the browser to steal information, all whereas bypassing Perplexity’s information protections utilizing trivial Base64-encoding methods. The assault doesn’t embrace any credential theft part as a result of the browser already has licensed entry to Gmail, Calendar, and different related providers.
It takes place over 5 steps, activating when a sufferer clicks on a specifically crafted URL, both despatched in a phishing e-mail or current in an online web page. As an alternative of taking the person to the “supposed” vacation spot, the URL instructs the Comet browser’s AI to execute a hidden immediate that captures the person’s information from, say, Gmail, obfuscates it utilizing Base64-encoding, and transmits the data to an endpoint underneath the attacker’s management.
The crafted URL is a question string directed on the Comet AI browser, with the malicious instruction added utilizing the “assortment” parameter of the URL, inflicting the agent to seek the advice of its reminiscence slightly than carry out a dwell internet search.
Whereas Perplexity has labeled the findings as having “no safety influence,” they as soon as once more spotlight how AI-native instruments introduce new safety dangers that may get round conventional defenses, enable dangerous actors to commandeer them to do their bidding, and expose customers and organizations to potential information theft within the course of.
In August 2020, Guardio Labs disclosed an assault method dubbed Scamlexity whereby browsers like Comet might be tricked by menace actors into interacting with phishing touchdown pages or counterfeit e-commerce storefronts with out the human person’s data or intervention.
“AI browsers are the following enterprise battleground,” Or Eshed, CEO of LayerX, mentioned. “When an attacker can direct your assistant with a hyperlink, the browser turns into a command-and-control level inside the corporate perimeter. Organizations should urgently consider controls that detect and neutralize malicious agent prompts earlier than these PoCs turn into widespread campaigns.”
