Singapore authorities warn that personal AI agents carry privacy and security risks and could be abused to access accounts, reveal sensitive information, or perform unauthorized actions if misused.

A joint advisory by the Cyber Security Agency of Singapore, GovTech, and the Infocomm Media Development Authority notes that personal AI agents can act on behalf of users and may access email, calendars, contacts, files, and more.

Personal AI agents can read, process, and share information from connected accounts, may operate in the background, and can be tricked by malicious actors, so users should limit permissions and retain control over actions.

Recommendations include blocking lateral movement in networks, strengthening multi-factor authentication, minimizing permission scopes and durations, and restricting irreversible actions unless human-approved.

Meta defended Muse as safe, noting it runs on an isolated virtual machine and advised users to review permissions and grant only necessary access.

Outside concerns include reports of Muse accessing messages and Amazon blocking Muse from shopping on its platform due to security and user-experience concerns.

Experts urge common standards across agents from different companies to ensure consistent understanding of permissions, purposes, and data handling when agents cooperate.

Alex Stamos criticized Muse AI for potentially accessing sensitive data like passwords and financial information, calling it totally insane.

Precautionary steps include limiting what the agent can see, using strong authentication, downloading only from official sources, stopping the agent and disconnecting accounts if something goes wrong, and preserving activity records.

Second precaution: retain control of actions with explicit approvals for consequential tasks, using one-time approvals and regularly reviewing permissions and activity logs.

Key precaution: connect only necessary apps and accounts, avoid banking/government/healthcare access, and ensure you do not share usernames, passwords, or one-time verification codes with AI agents.

Users may be liable for actions taken by agents on their behalf depending on circumstances and applicable law.