On Wednesday (28 Oct), Meta launched Muse, a personal artificial intelligence (AI) agent, in Singapore.
The agent connects to partners including Singapore Airlines, Shopee, Grab, Agoda, Klook, Lazada, ShopBack and Trip.com.
According to a press release seen by MS News on 28 Oct, Muse is designed to carry out tasks on a user’s behalf.
These include managing emails, coordinating schedules and making purchases.
Source: Meta website, for illustration purposes only
The launch comes as AI agents draw more scrutiny worldwide over how much access they should have to people’s personal data and accounts.
Muse connects with apps that many Singaporeans already use daily.
Its first local partners are Singapore Airlines (SIA), Shopee, Lazada, Grab, Agoda, Klook, Trip.com, and ShopBack.
Source: Image courtesy of Meta
Through these partners, users can ask Muse to research products and compare prices on Shopee or Lazada, plan grocery runs on GrabMart, or book a table through Grab Dine Out.
It can also look up flights, hotels, attraction tickets and reservations through SIA, Agoda, Klook, and Trip.com. ShopBack users can earn cashback and claim dining deals through the agent.
Source: Meta website, for illustration purposes only
However, not every feature will be ready from day one. Meta said some connector functions will roll out progressively after the launch.
Muse is available on iOS, Android and the muse.ai website. Meta said it will also come to AI glasses soon.
Most features are free, with paid subscription plans for those who want to do more with the agent.
Source: Instagram
Meta said it spent most of its effort on the project making Muse safer to use.
Users decide which apps and data Muse can access. The agent must also get approval before taking sensitive actions, such as sending an email or paying for something.
These approval requests appear as separate pop-ups in the app, not within the chat itself. According to Meta, this prevents the agent from approving actions on its own.
Each user’s Muse runs on a private cloud computer, which Meta calls the Muse Secure VM.
Passwords, login tokens and payment details are stored there, but kept hidden from the agent itself.
Source: Meta website
When shopping online, users get a confirmation prompt showing the exact purchase details every time.
For new merchants, Muse issues a single-use card number instead of passing on the user’s actual credit card details. The number is tied to one merchant and one amount, and expires after a short period.
Meta added that Muse’s email connector filters out one-time passwords, password reset links and login links. This is meant to stop the agent, or anyone manipulating it, from using a person’s inbox to break into their other accounts.
Despite these measures, Meta acknowledged that Muse “will sometimes make mistakes”.
One major risk is prompt injection. This is when hackers hide malicious instructions in emails, websites or images to trick an AI into doing something its user never asked for, such as leaking personal data.
Meta said this remains an unsolved problem across the industry. To guard against it, the company uses several layers of protection, including training its AI model to spot such attacks and running separate screening systems that the agent cannot switch off.
The company is also inviting the public to hunt for weaknesses in Muse. Its bug bounty programme pays up to US$300,000 (about S$380,000) for valid reports, including up to US$130,000 (about S$170,000) for successful prompt injection attacks affecting one user.
Source: Meta Bug Bounty website
Muse’s arrival follows a recent reminder from Singapore’s Personal Data Protection Commission (PDPC) about the data risks of AI tools.
In an advisory published on 30 Sept, the PDPC warned organisations against relying on AI-generated outputs without proper human oversight or testing.
It cited a scenario in which an employee used a public AI tool to create an email script, which then exposed the email addresses of all customers on a mailing list.
The commission urged organisations to keep track of the AI tools their staff use, limit access to personal data, and ensure humans review AI work before it goes live.
Also read: Driver allegedly falls asleep on expressway in China while using ‘smart-driving’ mode
Have news you must share? Get in touch with us via email at news@mustsharenews.com.
Featured image courtesy of Meta website and courtesy of Meta. Picture on the left for illustration purposes only.