Is your AI browser safe?

THESE SYSTEMS PROMISE SMART AUTOMATION BUT POSE SECURITY RISKS AS WELL

- SUDHIR CHOWDHARY

IF YOU'VE USED a modern browser lately, you’ve probably noticed something changing. These systems are no longer passive tools for opening tabs or trying in search bars; they’re turning into active, intelligent companions. An AI browser like Perplexity’s Comet, OpenAI Chat-GPT Atlas or Opera Aria, can understand what you want, find answers, summarise web pages, and even perform actions on your behalf. It’s like having a personal AI assistant that helps with everyday tasks.

While AI browsers unlock big productivity gains, they also introduce new and unseen security risks. Unlike conventional browsers that simply display content, AI-powered browsers actively interpret, summarise, autofill, and even execute tasks on behalf of users. This added layer of autonomy introduces new and unpredictable risks for both individuals and enterprises. Attackers target this “thinking” layer, using advanced tactics like prompt injection to influence browser behaviour.

A typical manipulation chain looks like this: An employee uses an AI browser to research or interact with a site while logged into corporate accounts. The agent fetches page content (including hidden or adversarial instructions) or accepts a screenshot/image. The adversarial content contains instructions that the model interprets as user intent. Attackers can deliver the malicious content via compromised web pages, or via malicious browser extensions/sidebars. Various audits have shown these flows can expose credentials, tokens, and confidential data.