Meta introduced a new AI agent called Muse on September 8. The agent is built to complete tasks for users instead of just answering questions.
Muse can send emails, book travel, create shopping lists, and make purchases on a person's behalf. Users interact with it through a dedicated app, a website, or WhatsApp.
The company says Muse runs on Meta Spark, its most capable model for this type of task based work. Meta built the tool through its Superintelligence Labs division.
How Muse Works
Muse operates inside something Meta calls a Muse Secure VM. This is a separate virtual computer in the cloud that holds a user's data and account credentials.
A second system called Sentinel sits alongside Muse. Sentinel decides whether Muse is allowed to take an action on the internet, and it can ask the user for approval first.
Meta says Muse does not see a user's passwords or payment details directly. Instead, sensitive information is stored separately so the agent can use it without viewing it.
For purchases, Muse can work with Stripe's Link system. This generates a one time use card number so a person's actual payment details stay hidden during checkout.
Users choose which apps Muse can connect to and how much access it receives. People can also disconnect a service or ask Muse to forget specific details it has learned.
Meta plans to release a version called Muse Confidential VM later this year. That version would be encrypted with a key only the user holds, meaning Meta itself could not view the contents.
Muse is rolling out first to adults in the United States on iOS, Android, and the web. Meta also says the agent will come to its AI glasses.
The service is free for most users, with paid plans available for people who want higher usage limits.
US Accuses Chinese Firms Over AI Models
On the same day, US officials accused six China based AI companies of extracting capabilities from American AI systems. The companies named include DeepSeek, Moonshot AI, and Alibaba.
The accusation centers on a method called distillation. This involves using the outputs of one AI model to help train another model.
The NSA, FBI, and CISA issued a joint advisory describing the activity as industrial scale. They said it could let Chinese developers close the gap with US firms without paying the full cost of building their own frontier models.
Officials said the targeted models included systems built by Anthropic, OpenAI, and Google. The advisory also said the activity was likely known to the Chinese government.
US officials said the technology gained through this method could affect China's military and cyberattack capabilities. The Chinese embassy in Washington had not responded to the claims as of the report.
The allegations follow earlier reports in July that Chinese military researchers had used outputs from US AI systems to train their own models. Similar claims were also made by US officials earlier this year.
The claims come as the US and China prepare for talks on AI safety. President Donald Trump and Chinese President Xi Jinping are expected to meet later in September.
