Researchers Trick Perplexity's Comet AI Browser Into Phishing Scam in Under Four Minutes

Researchers demonstrate how Perplexity's Comet AI browser can be manipulated into executing phishing scams within four minutes using a GAN-based attack.

• •The attack exploits "Agentic Blabbering," where the AI browser verbosely narrates its reasoning, exposing what it considers suspicious or safe to external observers.
• •By intercepting traffic between the browser and AI servers and using it as GAN training input, attackers can iteratively optimize a phishing page until the AI stops flagging it.
• •Once a phishing page bypasses one AI browser instance, it works against all users relying on the same model, shifting the attack target from humans to the AI itself.
• •Trail of Bits also demonstrated four prompt injection techniques against Comet to exfiltrate private data from Gmail.
• •