By Hamza Faisal

Singapore has issued a blunt warning to anyone handing their digital life to a personal AI agent: treat it like a helpful stranger, not a trusted assistant. The country's Cyber Security Agency (CSA), Government Technology Agency (GovTech) and Infocomm Media Development Authority (IMDA) jointly warned consumers on Friday that AI agents with access to their accounts can be manipulated into transferring money, revealing private information or taking actions they never asked for.

Unlike ordinary chatbots that simply answer questions, personal AI agents act on a user's behalf. They can browse websites, manage email, make bookings and log into online accounts — and, once connected, they can read, process and share information from those services, including personal data belonging to other people.

Helpful, but not to be trusted blindly

"AI agents can be very useful, but you cannot treat them like human assistants and give them the same broad access to your information and implicit trust," the advisory said. "Instead, grant narrowly scoped permissions, retain control of consequential actions and understand how to revoke access."

The agencies named several agents now in use, including Dots, Hermes, Meta's Muse, Grok Bot and OpenClaw. The warning follows the launch of Meta's Muse in Singapore on October 28, billed as the world's first personal AI agent for consumers, with links to services including Singapore Airlines, Shopee, Grab and Lazada.

Scammers are already adapting to the new tool. The advisory warned that hidden instructions planted in websites or emails can trick an agent into handing over sensitive data or moving money — with the user left holding the consequences. Depending on the circumstances, users may be held liable for actions their agents take on their behalf.

Five rules for safer agent use

The advisory sets out five precautions. First, connect only the apps and accounts genuinely needed for a specific task, and restrict the information each one can reach. Banking, government and healthcare accounts should stay off limits entirely, and users should never share passwords, one-time passwords, credit card details or Singpass credentials with an agent.

Second, keep control of anything consequential. Start with limited or read-only permissions where possible, and require the agent to ask before sending messages, deleting files, making purchases or posting publicly. One-time approval is safer than a standing "always allow" setting.

The remaining guidance covers checking how providers store data obtained through connected accounts, guarding against malicious instructions hidden in content the agent reads, and regularly reviewing and revoking permissions that are no longer needed. Workers were also advised to check with their employers before connecting work accounts.

A warning the rest of the world will likely copy

Singapore is among the first governments to issue formal consumer guidance specifically for personal AI agents, and the timing is deliberate: the tools are moving fast from novelty to everyday use, and the gap between what they can do and what users understand about them is widening.

The message is not to stop using AI agents, but to stop treating them as people. A human assistant can be trusted with context and judgment; an agent, the agencies stressed, follows instructions — including instructions planted by someone else. Until the technology earns that trust, the safest approach is the narrowest one: minimum access, explicit approval, and nothing sensitive within reach.