Personal AI agents from Meta and OpenAI are being offered for email and purchasing tasks, raising questions about how users control their data and spending when software acts for them.
Walmart said customers who use its Sparky shopping agent spend 40% more per order than customers who do not. The comparison does not show that Sparky caused the difference. Consumer protection expert Lindsay Owens told CBS News she is concerned that shopping agents could encourage people to spend more.
Meta began rolling out Muse in the United States on Sept. 8. The company says users can connect services to let the agent send emails, arrange travel and make purchases. OpenAI announced dots on Sept. 29 and began making it available in eligible markets. It says dots can use connected apps and operate on its own cloud computer.
Both companies describe controls on what their agents can do. Meta says a separate Sentinel agent reviews Muse’s internet activity, while Muse seeks a user’s approval before sensitive actions, including purchases. OpenAI says the tools used for dots’ proactive research are read-only and that users can set access and approval rules. The company also cautions that its agents can make mistakes.
Access creates another risk: instructions planted in material an agent reads could steer it toward revealing data or taking an action the user did not authorize. CBS News described how such instructions can appear in websites, emails and documents. Lynette Owens, a consumer education expert at TrendLife, warned CBS about malicious website material and prompt-injection incidents. Cybersecurity expert Hed Kovetz told CBS he would not give commercial agents access to his email, banking details or health records.
Users may also find practical savings. Chris Franco, founder of marketing firm Woodridge Growth, told CBS News that agents helped him identify unused subscriptions by sorting receipts in his Gmail account. Matt Fredrikson, a computer engineering expert at Carnegie Mellon University and founder of Gray Swan AI, told CBS News that users could ask an agent for a plain-language summary of how it completed a task.
