Agentic browser fell for phishing attack
Never happened before and here we go again. The Comet browser from Perplexity with agentic AI turned out to be catastrophically vulnerable to cyberattacks. But Guardio research revealed serious vulnerabilities. In one test they created a fake Walmart site through Lovable service. Asked Comet to buy an Apple Watch. The browser didn’t verify the site’s legitimacy, automatically filled credit card data and address, completed the purchase without a single user confirmation!
The second test was even worse. A fake Wells Fargo email from a ProtonMail address contained a link to an active phishing page. Comet perceived this as a genuine bank instruction, followed the link and suggested the user enter credentials on the phishing site!
The climax — attack through malicious prompts. Researchers created a fake CAPTCHA page with hidden instructions in source code. Comet interpreted them as valid commands, clicked the button and launched download of a malicious file!
But the scariest is yet to come. I quote Guardio — in the era of AI vs AI confrontation, fraudsters don’t need to deceive millions of people, it’s enough to hack one model. After success, the exploit scales infinitely. Hackers have access to the same models — they can train malicious versions against victim AI until fraud works flawlessly!
Are we on the brink of technological catastrophe? Agentic browsers promise a convenience revolution but open doors for new generation attacks. The question isn’t whether mass hacking will happen — the question is when.
