A Guth Labs publication ↗
Back to News

Singapore urges narrow permissions and human approval for personal AI agents

AI-generated conceptual illustration of an assistant connected to gated email, calendar and files, with a hand controlling access.
guthlabs.ai · AI-generated conceptual editorial illustration

Singapore's Cyber Security Agency, Government Technology Agency and Infocomm Media Development Authority issued joint guidance on November 6 for people using personal AI agents. The advisory focuses on assistants that can connect to accounts and act on a user's behalf, rather than simply generate conversational answers. Joint advisory announcement

The agencies recommend limiting access to what a task needs, retaining approval over consequential actions and knowing how to revoke access. They warn that malicious content an agent encounters can redirect its behaviour. The document provides consumer safety guidance; it does not announce a general prohibition on personal agents. Consumer guidance

Permission settings matter more than a reassuring conversation

The advisory recommends starting with limited or read-only permissions and using application settings rather than relying entirely on chat instructions. It favours one-time approval for important actions and regular checks of both access permissions and activity history. It also recommends avoiding sensitive account connections and keeping authentication credentials outside an agent's reach. Access and approval recommendations

The distinction is practical. A promise in a conversation describes intended behaviour. A permission restriction determines which operations the connected service will accept. Users evaluating a product should be able to identify both: what the agent says it will do, and what the underlying account settings actually allow.

Earlier guidance addressed deployment risks

The November advice builds on earlier Singapore work. A May 28 CSA advisory concerning OpenClaw identified risks including weak access controls, malicious extensions, exposed data and poisoned memory. It recommended least-privileged accounts and stronger controls for organizations using agents in sensitive environments. That earlier document is background, not a second new release this week. May cybersecurity advisory

For organizations, the May guidance also recommends logging actions, testing approval gates and rebuilding a compromised environment from a known-good baseline rather than assuming a restart solves the problem. It stresses that controls should be tested, including scenarios in which an agent must be prevented from acting. Organizational precautions

Plan for disconnection before something goes wrong

The new consumer advisory asks users to understand data collection, sharing, retention and what happens after an account is disconnected. If behaviour becomes suspicious, it advises stopping the agent, disconnecting affected services, reviewing unexpected changes and preserving activity records before cleanup. Data handling and response

For a prospective user, that creates a useful buying checklist. Is the stop control easy to find? Can access be removed from the connected account as well as the agent app? Does the history make consequential actions understandable? These questions can be asked before granting broad access.

The broader message is that convenience depends on deliberate configuration. Starting with a narrow, low-risk task gives users a chance to inspect the assistant's behaviour and controls before deciding whether a larger scope is justified.

Sources and citations

Revision history

Published .