AI Agents Tricked Into Launching Phishing Attacks in Under 4 Minutes

Security researchers demonstrated a concerning attack: they manipulated Perplexity's Comet AI browser to launch a complete phishing attack in under 4 minutes, without any human intervention. The exploit abuses a fundamental characteristic of browser-based AI agents: they need to process and interact with web content to function. Through prompt injection embedded in HTML, the agent was convinced to create a convincing phishing form to capture login credentials. This demonstrates that even products from reputable companies can be turned into attack tools if the underlying model lacks adequate protections against prompt injection. Implementers should enforce strict sandboxing, action allowlists, and human approval for sensitive actions.