Singapore authorities warn of risks of personal AI agents — Channel NewsAsia

Singaporean authorities have cautioned users of personal AI agents against granting such services broad access to applications and online accounts. The joint advisory was issued on November 6 by the Cyber Security Agency of Singapore (CSA), GovTech and the Infocomm Media Development Authority (IMDA), Channel NewsAsia reports.

The agencies noted that, unlike regular chatbots, personal AI agents can act on behalf of a user. They can connect to email, calendars, contacts, files and other services, browse websites, manage emails and make reservations. Examples of such agents included Dots, Hermes, Muse, Grok Bot and OpenClaw.

Access only to necessary services

Depending on the permissions granted, an AI agent may access and disclose sensitive personal information, make mistakes or perform actions unwanted by the user, the agencies warned. Some agents may also operate in the background, while malicious actors may attempt to mislead them with harmful instructions hidden on websites, in emails or messages.

More current news is available on the UA.News Telegram channel Telegram.

Users are advised to connect only the applications and accounts genuinely needed for a specific task and, where possible, further restrict access within each service. They are advised not to grant agents access to banking, government or medical accounts, or provide them with credit card details, passwords, access codes or one-time verification codes.

Control over important actions

The agencies recommend starting with limited permissions or read-only modes and requiring separate confirmation before actions that may have significant consequences. These include sending messages, deleting files, publishing content or making purchases. Permissions should be changed in application settings rather than relying solely on commands given to the agent in a chat.

Users are also advised to regularly review granted access and credentials, revoke unnecessary permissions and monitor agents' activity. If a service behaves unexpectedly or there is suspicion that information has been disclosed, users should stop the current task, disconnect linked accounts and check for changes they did not make. The authorities stressed that, depending on the circumstances and applicable law, users may be held liable for actions carried out by agents on their behalf.