Meta’s Muse AI agent: What can it do and what precautions should users take?

SINGAPORE - From Oct 28, Singapore users can download Meta’s Muse from the Apple App Store or Google Play Store, establishing Muse as the first artificial intelligence agent designed for the mainstream consumer market.

Muse’s rollout here follows its global debut in the United States and Canada on Sept 8, and a subsequent launch in Japan on Oct 23. It has surpassed 10 million downloads within just 40 days of its debut, and became the most downloaded free app on the Apple App Store here within hours of its launch in Singapore.

Here’s what you need to know about Muse and other AI assistants before using them.

1. What can Muse do, and how does it fare against other free AI tools?

Users need to give Muse permission to access data in Gmail, Instagram and Calendar for the AI assistant to execute tasks on their behalf or send reminders for appointments. Users can also connect other AI tools such as OpenAI’s ChatGPT, Anthropic’s Claude, and Google’s Gemini to everyday apps such as Gmail and Google Calendar.

A test done by The Straits Times found that the extent of this autonomous action varies across the free tiers of these AI tools.

Basic tasks such as drafting and sending an e-mail could be accomplished by all four AI tools. On the retail front, all four tools were able to recommend a list of coloured contact lenses sold in Singapore, but only Muse was able to accurately provide the real-time availability of specific prescriptions and complete the purchase.

Asked to plan a night out with friends, all four AI tools were able to recommend a schedule that starts with a movie followed by dinner places and bars within walking distance. But Meta’s Muse took further steps by scraping live showtimes across multiple cinemas, checking real-time seat availability and calculating the total price for the group for different ticketing options.

Gemini was prone to hallucinations, providing inaccurate show timings and contact lens availability. Muse was eager to assist even when not instructed. With access to the iPhone calendar, Muse was able to pick up on an upcoming overseas trip, and sent a unprompted reminder to complete the digital arrival card and apply for a visa for business travel.

Muse is free to use with a weekly usage limit, which Meta said is sufficient for most of what people need. Those looking to do more complex work can pay between $22 and $139 for subscription plans which come with more Muse tokens or credits.

But the autonomous features of ChatGPT, Claude, and Gemini are mostly locked behind paid subscription plans.

2. What are the risks that AI agents such as Muse pose?

Singapore users will be able to progressively connect Muse to apps including Singapore Airlines, Shopee, Lazada, Grab, ShopBack, Agoda, Klook and Trip.com.

Connecting Muse to these apps will allow the agent to identify the most competitive prices of items and flight tickets, order the groceries needed for a specific meal, and book travel experiences based on one’s preferences and holiday plans.

To allow Muse to make payment, users will need to save their card details on third-party digital wallet Link by Stripe. The wallet will generate a unique one-time card number for each purchase, so that real card details stay private.

Yifan Jia, founder of AI safety testing firm AIDX Tech, said one risk is prompt injection. It is a technique where attackers embed malicious instructions in content that AI agents read, such as websites, documents and e-mails.

“An attacker could send an e-mail containing hidden instructions telling the agent to retrieve confidential documents and forward them to an external address,” said Jia. “An AI agent tasked to organise a user’s inbox might fail to distinguish legitimate user instructions from untrusted content, and could unintentionally execute the attack.”

Another risk is automation bias, where users may gradually stop checking the actions of AI agents, similar to how many people have stopped checking automated deductions on their bank statements.

“For mass market adoption, providers have a particularly important responsibility, (as) most consumers are unlikely to understand complex security settings or adjust default permissions,” said Jia, adding that the onus is on providers to build safety into the product by default.

3. What precautions should users take when using AI agents?

Access to external apps and documents should only be granted according to what each task requires, said April Chin, co-founder of AI safety testing firm Resaro. For instance, an agent that manages a user’s calendar might not require access to one’s e-mail inbox.

“Connect one app at a time, and hold off on sensitive data such as banking and health data until you are clear about the scope of activities the agent is allowed and not allowed to do,” said Chin.

Generally, users can provide AI assistants with read-only access to external apps. This prevents AI assistants from sending messages, modifying files or making purchases.

Chin also urged users to check the activity log of AI assistants regularly and revoke access that is no longer needed.

Actions that cannot be undone such as payments and outgoing messages should require human approvals, said Jia, adding that users should avoid unnecessarily sharing passwords, Singpass credentials, NRIC numbers or financial information.

Although Muse seeks approval from users before making payment or sending an e-mail, Jia cautioned against granting AI assistants additional system privileges. “Be cautious when an agent requests additional system privileges and understand why those permissions are needed.”

4. Should I connect AI agents to my work e-mail, systems?

Personal AI agents should not be connected to one’s work e-mail or documents without explicit approval from management, said Chin, adding that this can create a blind spot that IT and security teams cannot see or control.

A compromised agent linked to work accounts can also give attackers a direct route into corporate systems, said Jia. She added that employers should establish clear AI usage policies, implement access controls and monitoring mechanisms, conduct security assessments, and provide employees with approved AI tools.

“Staff will adopt these tools with or without a policy - so the real choice is between managed and unmanaged use,” said Jia.