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

The report describes how agentic AI browsers — which autonomously act across websites on behalf of users — can be tricked through prompt injection and adversarial training (e.g., using GANs) into performing malicious actions like submitting credentials or leaking data. Researchers demonstrated making Perplexity's Comet fall for a phishing scam in under four minutes and other groups disclosed prompt-injection and zero-click techniques that extract private information, highlighting a growing attack surface where adversaries train scams against the AI agent rather than the human user.