The two developments highlight the growing competition around AI agents and the broader race to develop increasingly capable artificial intelligence systems. Meta is positioning Muse around privacy and security, while U.S. officials are raising concerns about the use of AI model outputs by Chinese companies.
Meta Muse Brings Personal AI Agents to Everyday Tasks
Meta introduced Muse on September 8 as a personal AI agent that can move beyond answering questions and take actions for users. Initially available to adults in the U.S., the service can be accessed through a dedicated app, the web, and WhatsApp. Meta also plans to bring Muse to its AI glasses.
Meta launched Muse, a proactive AI agent for appointments, shopping, travel planning, and goal management. Source: @Muse via X
Unlike conventional chatbots, Muse is designed to operate across connected applications and complete tasks with less step-by-step guidance. Meta says users can ask it to send emails, organize plans, book travel, create shopping lists, and make purchases.
The agent is powered by Muse Spark, which Meta describes as its most capable model for agentic tasks. The company says the system is designed so users can simply describe what they want done rather than learn a complicated workflow.
Muse was developed by Meta Superintelligence Labs, the company’s AI division established as part of its broader effort to compete with companies including OpenAI and Anthropic. The launch also reflects Chief Executive Mark Zuckerberg’s strategy of making personal AI agents a central part of Meta’s consumer technology ecosystem.
The company says Muse can remember information that users choose to share and use it to make suggestions. For example, it can turn a recipe saved on Instagram into a grocery list or help organize a dinner based on previously provided preferences.
Meta is also integrating Muse with third-party services. Its capabilities can extend to email, calendars, shopping, payments, fitness, smart-home systems, and other applications, depending on the permissions granted by the user.
For payments, Meta says Muse can work with Stripe’s Link infrastructure, which is designed to provide an agent with a single-use card number instead of exposing a user’s underlying payment credentials. Meta says this approach is intended to reduce the risks associated with allowing an AI system to make purchases.
Secure by design, sort of
Security is one of Meta’s central selling points for Muse because the agent may require access to significantly more sensitive information than a conventional chatbot.
Meta says Muse operates inside a dedicated Muse Secure VM, a virtual machine in the company’s cloud that contains the user’s data and credentials. The environment includes its own browser and is isolated from other users’ agents.
Zuckerberg says no competing AI agent matches Muse’s secure, self-configuring cloud VM setup. Source: Rohan Paul via X
A separate system called Sentinel acts as the permission authority between Muse and the outside internet. According to Meta, Muse can propose an action, but Sentinel determines whether it can proceed based on policies established by the user. Sensitive actions can require explicit approval.
“The agent proposes actions, but only Sentinel can grant permission to perform an action,” Meta’s technical documentation says.
Meta also says Muse does not directly see users’ passwords or payment information. Users can decide which applications the agent can access and what permissions it receives. They can subsequently change those permissions or disconnect services.
The company says Muse keeps an audit trail of its actions and does not share conversations or data stored inside the user’s virtual machine with Meta’s advertising systems. Users can also opt out of having their interactions used to train Meta’s AI models.
The security model nevertheless leaves an important question around how much users must ultimately trust Meta itself. The company has acknowledged that its initial Secure VM architecture does not provide the same level of cryptographic isolation as a system in which the user alone controls the encryption keys.
Meta says it plans to introduce Muse Confidential VM later this year. Under that system, the virtual machine and its data would be encrypted with a key controlled by the user, meaning Meta would not be able to access the contents of the environment.
The issue is particularly relevant as AI agents become capable of taking real-world actions. Giving an agent access to email, financial services, calendars, and other personal systems creates a different security challenge from using a chatbot simply to generate text.
Meta’s launch comes amid broader scrutiny of autonomous AI systems, including concerns about prompt injection, unauthorized actions, and the potential for agents to interact with external services in unexpected ways. Meta says it has built Muse with these risks in mind and has included safeguards intended to restrict external activity.
Muse is initially free for most users, while Meta is also offering paid options for people seeking higher usage. The company says the service is rolling out in the U.S. across iOS, Android, and the web, with support for AI glasses planned.
U.S. accuses Chinese AI firms of maliciously copying American technology
The Muse launch comes on the same day that U.S. officials accused six China-based AI companies of conducting what they described as industrial-scale efforts to extract capabilities from American AI models.
The U.S. government accuses Chinese AI firms of industrial-scale “distillation” using outputs from leading U.S. AI systems. Source: Reuters via X
The companies named by U.S. officials include DeepSeek, Moonshot AI and Alibaba. The allegations center on a technique known as AI model distillation, in which the outputs of a larger model are used to help train another model.
Distillation itself is a legitimate technique used throughout AI research. However, the U.S. government alleges that the companies targeted proprietary capabilities from leading American AI systems in ways that allowed them to reduce the cost and time required to develop competing models.
In a joint cybersecurity advisory, the NSA, FBI and CISA said China-based companies were conducting “aggressive, industrial-scale distillation activities” against U.S. frontier AI models. The agencies said the activity could allow Chinese developers to narrow the technology gap without bearing the same research, computing, and infrastructure costs associated with developing frontier models independently.
U.S. officials said the companies had targeted models developed by American AI firms, including Anthropic, OpenAI, and Google’s AI operations. The Reuters report also said officials alleged the activity was likely carried out with Chinese government awareness.
The allegations extend beyond commercial competition. U.S. officials argued that access to advanced AI capabilities could have implications for China’s military and cyber capabilities.
“The technology could be used to advance China’s military and cyberattack capabilities,” officials said in their statement, according to Reuters.
The Chinese embassy in Washington had not immediately responded to requests for comment on the allegations, Reuters reported.
The dispute arrives at a sensitive point in U.S.-China relations. Washington and Beijing are preparing for discussions involving AI safety, while President Donald Trump and Chinese President Xi Jinping are also expected to meet later in September, according to Reuters.
The accusations are not entirely new. U.S. officials made similar claims earlier this year, while Reuters reported in July that Chinese military researchers had used outputs from leading U.S. AI systems to help train domestic AI models.
The latest allegations nevertheless underscore how quickly the AI race has moved beyond model performance. Access to advanced models, computing resources, proprietary data and the ability to deploy AI agents are increasingly becoming strategic issues for both technology companies and governments.
For Meta, Muse represents another step toward an AI model that can actively interact with the digital world rather than simply respond to prompts. For U.S. policymakers, the allegations against Chinese AI companies demonstrate the growing concern that access to advanced AI capabilities could influence both commercial competitiveness and national security.
The two developments point to the same broader shift: AI is increasingly moving from systems that generate information to systems capable of taking action. As that transition accelerates, security, data control, and the provenance of AI technology are becoming as important as the underlying models themselves.
