Meta’s ambitious expansion into personal artificial intelligence agents is facing an early test over whether autonomous software can be trusted with access to users’ private information and digital activities. Its Muse AI agent, introduced as a tool capable of managing everyday tasks and interacting with different applications, has reportedly generated messages containing information users never provided, accessed conversations from other messaging services and produced inaccurate details about financial transactions.
The incidents raise questions about the safeguards surrounding a technology that is designed to do considerably more than respond to instructions. Unlike conventional chatbots, autonomous agents are expected to interact with applications, retrieve information, make decisions and complete tasks on behalf of users. This broader access creates commercial opportunities for technology companies, but also introduces privacy and security risks that are considerably more difficult to manage.
Rapid Expansion Has Increased Pressure on Safety Systems
Muse has attracted more than 10 million downloads since its September launch, making it an important part of Meta’s strategy to expand its artificial intelligence business. The company wants to move beyond conventional chatbot services and develop an assistant that can understand users’ preferences, remember information and perform activities across different digital platforms.
The commercial opportunity is substantial. Meta already operates Facebook, Instagram and WhatsApp, giving it access to an enormous global user base. Integrating personal AI agents into this ecosystem could create new opportunities for subscriptions, digital services and automated commercial interactions.
However, the early problems reported by employees suggest that the development of autonomous capabilities may be creating difficulties that are not immediately apparent during ordinary testing. Internal company discussions reviewed in the investigation described incidents involving unexpected messages, inaccurate financial information and apparent references to conversations that belonged to other users.
In one reported case, an employee received information about bank account balances despite never having discussed banking activities with the agent. Another employee said Muse unexpectedly reminded her to respond to a message received through Signal, an encrypted messaging application.
These incidents have raised concerns about whether the agent could be accessing information beyond what users believe they have authorised. Meta has described several cases as isolated incidents or hallucinations and said it is continuously improving the system’s safeguards.
Nevertheless, the reported problems raise a more fundamental question about how much access an AI assistant should receive before users can reasonably trust it with sensitive information.
Broad Device Permissions Create Unexpected Risks
One of the more significant aspects of the investigation concerns the way Muse reportedly accessed information from applications outside its direct connections. Tests conducted by Reuters with assistance from privacy specialists indicated that the agent could retrieve message content through permissions granted during its initial setup.
On Android devices, the agent was able to use access to push notifications to retrieve information from other applications. On Apple computers, investigators found that accessibility and screenshot permissions could provide another route for obtaining information. The same behaviour was not reproduced on iPhones during the reported tests.
The findings highlight a weakness in the way users understand digital permissions. A person may grant an application access to notifications or accessibility features for a particular purpose without realising that an AI agent could use those permissions to retrieve information from other services.
This becomes especially important when an application is designed to operate independently and perform tasks without requiring constant instructions. A traditional application generally performs a relatively defined set of functions, whereas an autonomous agent may interpret information from different sources to complete a broader objective.
The distinction creates a new security problem. Even when permissions are technically granted by users, the way those permissions are subsequently used may extend beyond what people reasonably expect.
Privacy protections therefore cannot depend entirely on asking users to approve broad access during installation. Technology companies must also explain how information will be collected, interpreted, stored and used during subsequent interactions.
Incorrect Information Can Become a Financial Liability
Privacy concerns are not the only problem associated with Muse. Internal employee discussions also reportedly described cases in which the agent supplied inaccurate information about purchases, product prices and financial activities.
In one incident, an employee reported that Muse had booked a non-refundable accommodation reservation without permission. Other reported cases involved the agent referring to purchases that had not occurred or providing incorrect information about transactions.
Such incidents demonstrate why autonomous AI systems require stronger safeguards than conventional information tools. A chatbot providing an inaccurate answer may mislead a user, but an agent capable of taking action can turn an incorrect interpretation into a financial or administrative problem.
The distinction becomes particularly important when an agent is connected to shopping platforms, travel services, email accounts or financial applications. A mistake in understanding an instruction could result in an unwanted purchase, an incorrect booking or the disclosure of confidential information.
Meta has said that Muse requires users to provide permissions before accessing connected applications and that sensitive activities should be subject to safeguards. Its published product design also describes a separate security system intended to review actions before they reach the internet.
However, the reported incidents suggest that the practical implementation of these protections needs to be examined alongside the company’s stated security architecture.
The problem is not necessarily that autonomous agents cannot be made safe. It is that the consequences of mistakes become more serious as their ability to interact with external services increases.
AI Hallucinations Are Becoming More Complicated
Artificial intelligence hallucinations are not new. Generative AI systems can produce information that appears convincing but is factually incorrect, particularly when asked to answer questions involving incomplete or ambiguous information.
However, the reported Muse incidents suggest that hallucinations in autonomous systems can create additional complications. Some employees described receiving unexpected messages involving childcare, financial information and purchases that were unrelated to their conversations.
Meta employees reportedly suggested that built-in prompts designed to encourage engagement between user sessions could be contributing to some of these incidents. The company has also acknowledged that problems may involve both the underlying AI model and the software that enables it to perform actions.
This distinction is important because improving the language model alone may not resolve problems created by the surrounding software. An AI agent depends on several interconnected components, including permissions, memory, application connections, task management and systems that determine whether an action should be performed.
A failure in any of these components could result in unexpected behaviour. For example, an agent may correctly interpret an instruction but retrieve information from the wrong application, or it may generate an inaccurate response based on information that was never intended for that user.
Consequently, testing autonomous AI requires more than evaluating whether a model produces accurate answers. Developers must examine how the entire system behaves when multiple applications, accounts and information sources are involved.
Technology Companies Face Growing Accountability Questions
Meta’s experience also highlights a wider problem for the technology industry. Companies developing autonomous agents are competing to establish their products before the market becomes crowded, but the technology is evolving faster than many existing approaches to digital privacy and security.
OpenAI, Anthropic and other developers are also working on systems capable of interacting with external applications and completing tasks independently. These products could reduce the time users spend managing routine activities, but they also introduce vulnerabilities associated with excessive permissions, inaccurate instructions and unexpected interactions with external systems.
Some financial institutions and retailers have already begun restricting access by autonomous agents because of concerns about how these systems interact with their platforms. Operating system developers are also facing pressure to provide more precise controls over the permissions granted to AI applications.
Apple’s reported changes to its computer accessibility permissions demonstrate how the development of autonomous software could influence the design of operating systems themselves. Google and Apple have also indicated that they are examining ways to give users greater control over the information accessible to AI agents.
Signal has raised concerns about the use of device permissions to retrieve information from encrypted conversations. Its position highlights an important distinction: encryption can protect communications within a messaging service, but it does not necessarily prevent another application from accessing message content displayed on a device through operating system permissions.
This creates a difficult challenge for developers. Privacy protections must operate across different applications and devices rather than depending exclusively on the security arrangements of individual platforms.
Commercial Ambitions Must Not Outpace User Protection
For Meta, the success of Muse could become an important part of its future business strategy. The company has access to billions of users across its existing platforms and is seeking new ways to turn artificial intelligence into a major source of revenue.
However, the commercial potential of autonomous agents depends heavily on consumer confidence. Users may be willing to experiment with a chatbot that occasionally produces an inaccurate response, but they are less likely to accept a service that unexpectedly accesses private conversations or carries out financial transactions without clear approval.
Trust is particularly important because personal AI agents are intended to handle information that people may not willingly share with ordinary applications. Emails, financial details, appointments, shopping activities and private conversations could all become part of the information available to an assistant.
If users cannot understand exactly what the agent can access and when it can act, the convenience offered by the technology could be outweighed by concerns about privacy and control.
Meta will therefore need to demonstrate that its safeguards work consistently under real-world conditions. This includes clearer permission systems, stronger separation between user accounts, reliable approval mechanisms for financial activities and transparent records of what the agent has accessed or done.
The reported incidents do not establish that Muse routinely exposes one user’s private information to another. However, they demonstrate why such risks must be investigated before autonomous agents become deeply embedded in everyday digital activities.
The larger challenge facing the industry is no longer simply developing AI systems capable of completing complicated tasks. It is ensuring that these systems can do so without exceeding the authority users have given them.
For Meta and its competitors, the next stage of development will depend on whether autonomous AI can establish the same level of trust that consumers expect from the digital services they already use. Rapid adoption may create commercial momentum, but lasting success will require much stronger evidence that convenience does not come at the expense of privacy and security.
(Adapted from Fidelity.com)









