Security researchers recently found that some AI-powered browsers can be tricked into revealing saved passwords. This exploit, dubbed BioShocking, was discovered by cybersecurity firm LayerX. It takes advantage of a fundamental feature in AI browsers — their ability to read and respond to webpage content. The attack works because AI agents in these browsers follow instructions hidden within the content they analyze. Researchers call this a prompt injection attack, where malicious commands are disguised as normal content to hijack the AI’s actions. BioShocking is particularly clever because it disguises these harmful instructions within a game-like interface, potentially evading safety filters that look for obvious attack patterns. LayerX’s research showed that this tactic worked against multiple AI browsers, and at the time of their report, some vendors still hadn’t issued fixes. The research highlighted AI browsers in general, with Perplexity’s Comet mentioned as one example. The common thread among vulnerable browsers is that they have an AI agent capable of accessing stored credentials and acting on webpage content. The concerning part is that some vendors haven’t yet patched this vulnerability.