Should you trust personal AI agents? What security experts say.

Cybersecurity experts are divided on the safety of personal artificial intelligence agents, with some urging caution due to inadequate security measures while others argue the tools are viable if used with specific safeguards. As adoption rates surge, particularly among older demographics, questions persist regarding how much control users retain over these digital assistants. Hed Kovetz, a cybersecurity specialist, stated that he does not trust commercially available agents, such as Meta’s Muse or OpenAI’s Dots, with access to sensitive data including emails, banking details, and health records. He argued that major technology firms rushed these products to market without implementing sufficient security protocols. Despite acknowledging the commercial potential of such agents, Kovetz advised consumers to limit the amount of personal information they share. He noted that while some users claim they have nothing to hide, everyone possesses sensitive records such as medical files and financial accounts that remain vulnerable. Other specialists offer a more nuanced perspective, suggesting that consumers can utilize these tools effectively if they understand the inherent risks. Lynette Owens, a consumer education expert at the organization TrendLife, warned against granting AI agents excessive access or authority. She recommended that users allow agents to perform research tasks, such as comparing product prices, but cautioned against handing over purchasing power. Owens specifically advised against allowing agents to use credit card credentials. Her primary concern involves a technique known as prompt injection, where cybercriminals plant malicious instructions within webpages, emails, or documents. Unlike traditional attacks that require breaking into accounts, prompt injection tricks an AI agent into ignoring its original programming, such as limits on spending amounts. This vulnerability means an agent could be manipulated into taking unauthorized actions while executing a legitimate task. Consumer adoption of these technologies has been rapid. Meta’s Muse, launched in September, became the most downloaded product on Apple’s App Store shortly after its release. According to data from Sensor Tower, the application achieved ten million U.S. downloads within forty days, significantly outpacing ChatGPT, which recorded 3.7 million downloads in the same period. Muse currently boasts over 2.5 million daily active users and retains users at a higher rate than competing AI applications. The platform was designed to be beginner-friendly, featuring a customizable avatar to appeal to individuals unfamiliar with artificial intelligence. Alexandr Wang, Meta’s chief AI officer, described the product as being built for mothers, indicating a strategy targeted at non-technical users. Meta has actively promoted Muse through its existing platforms, including Instagram, Facebook, and WhatsApp. Demographic trends indicate that older adults are embracing these tools at a high rate. Apptopia estimated in September that 46 percent of Muse’s U.S. user base is aged 46 or older, making it skew older than any other AI chatbot. One user, 82-year-old Eileen Fay of Cranford, New Jersey, credited the agent with saving her thousands of dollars. Fay used Muse to analyze changes in federal import rules affecting her prescription medications from Canada. The agent generated a two-page fact sheet comparing Medicare Part D plans and identified an option that could reduce her annual costs by more than $6,000. Fay verified the accuracy of the information and noted that completing the analysis manually would have required extensive calculation. She now interacts with her agent, whom she named Siobhan, multiple times daily to manage calendars and research local issues. Privacy concerns accompany this growth. S. Shyam Sundar, director of Penn State’s Center for Socially Responsible Artificial Intelligence, expressed worry that the friendly design of these agents might lower the guard of older users. He noted that the technology has become sophisticated enough to maintain credible human-like interactions, which could lead users to disclose sensitive personal data inadvertently. Meta faces ongoing scrutiny regarding its data handling practices. In a separate legal development, a New Mexico judge ordered the company to pay $25 billion for violations of state consumer protection laws. Additionally, Meta’s AI glasses program has faced criticism over privacy risks. In March, following reports by Swedish publications Svenska Dagbladet and Göteborgs-Posten, the company was sued after it emerged that contractors in Kenya had accessed video footage from users of the devices. One worker reportedly stated that employees could view everything from living rooms to intimate scenes. Despite these controversies, Meta has expanded its accessibility initiatives. In June, the company announced that all legally blind veterans in the United States could receive its AI glasses free of charge. The Department of Veterans Affairs estimates that approximately 130,000 American veterans are legally blind, with over one million suffering from vision impairments that hinder daily activities. Don Overton, a blind veteran and former executive director of the Blinded Veterans Association, described regaining his independence upon using the Ray-Ban Meta glasses. Overton lost his sight in a bunker explosion and previously felt abandoned by the VA system. Historical investigations have highlighted systemic failures in supporting visually impaired veterans. A 2022 audit by the VA Office of Inspector General found that 87 percent of claims lacked evidence of required phone calls to visually impaired veterans. Furthermore, a 2022 Senate committee investigation determined that roughly 90 percent of VA websites did not meet federal accessibility standards. Consumer protection experts continue to monitor the behavior of these agents. Lindsay Owens, another consumer protection specialist, pointed out that it can be difficult for users to determine if an agent is acting in their interest or that of a retailer. She cited Walmart’s shopping agent, Sparky, which increased average consumer basket sizes by nearly 40 percent compared to standard browsing. This statistic suggests that some agents may be programmed to encourage higher spending. While the convenience offered by personal AI agents is evident, the lack of clear accountability mechanisms leaves users with limited recourse if an agent takes unauthorized actions. The debate centers on balancing the efficiency gains of automated assistance with the need for robust security frameworks that prevent exploitation and protect user autonomy.