AI Browsers: 4 of 7 Have Serious Security Flaws

A new study reveals that four of the seven most popular AI browsers have serious security flaws. These flaws could allow malicious websites to steal data from other sites you have open. Here's the thing: since 1995, browsers have used a rule called the "same-origin policy" to prevent websites from accessing each other's data. AI browsers bypass this rule to work across multiple tabs, and this broader access creates vulnerabilities. Attackers could use prompt injection to trick AI agents into revealing private information like emails or passwords. Another method, called memory poisoning, could plant instructions that activate later. Researchers successfully demonstrated this risk on ChatGPT Atlas. Claude for Chrome was flagged as particularly risky due to its design. The vulnerable browsers include ChatGPT Atlas, Chrome with Gemini, Claude for Chrome, and Perplexity Comet. Microsoft Edge with Copilot, Brave Leo, and Firefox AI Mode showed stronger security. Some companies, like Google and Microsoft, engaged constructively with the findings, while others, like Perplexity and OpenAI, declined to act. This suggests AI browser security may not be keeping pace with their rapid development.