How does Meta’s Muse fare against other free AI tools; what precautions should users take?

Meta's Muse, the first AI agent for mainstream consumers, has seen unprecedented success since its launch in Singapore on October 28. The AI agent, which can operate across various apps such as Gmail, Instagram, and Calendar, has already amassed over 10 million downloads in just four weeks. However, with the rise of such AI assistants, users must be aware of potential risks and take necessary precautions to ensure their safety.

One of the main concerns is prompt injection, a technique where attackers embed malicious instructions within content that AI agents read, such as websites, documents, and emails. For instance, an attacker could craft an email containing hidden instructions, prompting the AI to retrieve sensitive documents and forward them to an external address.

Another risk is automation bias, where users may become complacent and stop monitoring the actions of AI agents, similar to how individuals often overlook automated deductions on their bank statements.

To mitigate these risks, experts recommend granting Muse and other AI agents limited access to external apps and documents. For example, a calendar management agent may not need access to an email inbox. Additionally, users should regularly review the activity logs of their AI assistants and revoke any unnecessary permissions. When it comes to actions that cannot be undone, such as payments or message sending, users should always seek human approval and avoid sharing sensitive information like passwords, Singpass credentials, NRIC numbers, or financial data.