Commentary Outlines Privacy, Security and Reliability Risks in Meta's Muse AI Assistant

The commentary describes Muse as different from conversational tools such as Claude, ChatGPT and Gemini. It says Meta's platform is an autonomous personal assistant that manages tasks in the background. Meta is giving the AI deeper control over systems and apps, along with around-the-clock autonomous background execution, the commentary says. While the author likes the idea of an AI that works silently in the background, the article argues that intriguing ideas and reality often do not match. More autonomy means more chances for mistakes, security issues and other problems, according to the commentary.

The first concern is privacy and security. According to the commentary, Muse requires a high level of permission access, including access to personal and business accounts, financial details and email addresses. It also gives every user a virtual machine that is persistently available in the cloud. The author writes that combining a virtual machine with high-level data access is a recipe for potential problems. The commentary points to a zero-day macOS vulnerability in the Muse app that had to be fixed. The flaw allowed local malware to redirect dictation traffic and hijack the assistant's privileges. Even though the issue was addressed early, the author says it shows the risk is real. The commentary adds that it may only be a matter of time before a dedicated hacker, or even another AI, finds a way to exploit data from these virtual machines. The author acknowledges that Meta's approach could prove more secure than expected but says he would not count on it.

A second concern is hallucination. The commentary notes that AI tools are more accurate and powerful than before but remain far from perfect. Conversational bots still make mistakes and hallucinate, and the same underlying principles are used for agentic AI. The difference, the article says, is that agentic AI mostly works on its own to perform multi-step tasks. Muse does require approval once it works out a plan, at least in theory, but the commentary says that alone does not guarantee smooth sailing. It offers an example in which a user instructs an AI to collect data from a variety of sources for a business report, providing exact sources and step-by-step instructions. The AI returns an action plan, and the user signs off because everything looks good at a glance. Later, the user finds that much of the information was based on phantom data because the AI misinterpreted one source and tilted the results. By then, the report has already been sent to management. The commentary says such a mistake could lose a client or cost someone a job. It adds that professionals such as lawyers or teachers have been considered fully accountable when false AI data was used, and that the risk is not worth taking.

The third concern is that agentic AI can make major mistakes that are not easy to undo. The commentary says hallucinations in a conversational bot may confuse users but usually do not lead to major long-term financial problems or other serious repercussions because they are usually single mistakes. With an agentic system instructed to complete multiple steps autonomously, mistakes can compound. As an example, the commentary cites an earlier case this year in which a Claude-powered AI coding agent deleted an entire database in nine seconds and killed the backups in the process. It says that is far from the only example. With Muse, the commentary argues, users are not just giving the AI access to a few projects. Muse can connect to a user's entire Meta-based social apps, projects and finances. Some actions could even hurt a user's reputation without the user knowing, according to the article. Some AI mistakes are easy to spot, ignore and fix, but the commentary says Meta has the potential to do much more damage because of how much access it has over user data.

The commentary also mentions an alleged case of Meta Muse going rogue. It says a user asked Muse to handle a Facebook Marketplace listing, and the AI then handled the whole process. The article presents this as another example of the risk that comes with granting an autonomous assistant broad control.