Why your new AI assistant could be a security risk

Meta recently released Muse, an AI agent designed to manage your emails, book appointments, and shop online. At the same time, companies are rolling out smart glasses with cameras that look like everyday fashion accessories. Both technologies have run into trouble. Security researchers found that Muse could be tricked into handing control of a user's entire account to an attacker. Meanwhile, in India, smart glasses are being used to record people in public without their permission, leading to targeted online harassment and even police surveillance of protesters.

To be helpful, an AI agent needs deep access to your life—your calendar, your emails, and your payment details. Usually, your computer uses a digital key, called a token, to prove who you are when you log into a service. In the case of Muse, researchers discovered a vulnerability where a malicious program could hijack that token. Once the hacker has your token, the AI essentially acts as a puppet for them, following their commands rather than yours. Instead of the AI shopping for you, it could be used to download your private files or take photos using your computer’s camera without you knowing.

We are currently in a phase where companies are rushing to build powerful AI tools before they have fully secured the platforms they run on. When you give an AI assistant permission to send emails or buy goods, you are essentially granting a digital intern total access to your accounts. If that intern has a flaw in its design, the damage can be instant and widespread. We have to decide if the convenience of having an agent handle our errands is worth the risk of those same tools being used to watch us, or to let a stranger walk through our digital front door. The challenge is not just how to make these agents smart, but how to ensure they remain safely under our control.