AI agents promise privacy, but will they deliver?

At OpenAI's annual developer conference, chief executive Sam Altman unveiled Dots, a new AI agent, and declared the company aims to "set a new standard for privacy in advanced AI." Throughout the event, OpenAI speakers took thinly veiled shots at Meta's Muse, the main rival, over its alleged failure to protect user data.

The backdrop is Muse's launch a few months earlier as a safer alternative to its predecessor, OpenClaw. Mark Zuckerberg stated the product was "built from the ground up for privacy and security." Nat Friedman, head of product at Meta's superintelligence labs, wrote on X that the goal was to build a version of OpenClaw that could be made safe, convenient and scalable to billions of people. The architecture rests on a secure virtual machine — what Zuckerberg described as "an isolated Linux computer with a browser, CPU, memory and storage." According to the official blog, most development effort went into "careful design and engineering for safer operation," and the company acknowledged Muse would still make mistakes, but far less often and with limited damage thanks to its safety systems.

Despite reaching 600 thousand daily active users in the United States within weeks, per Apptopia, and topping the App Store charts, Meta's promises have not held up. Data is isolated between users, and Meta plans to introduce a cryptographic mechanism this year that would prevent even itself from accessing the VM. For now, however, the company can still read that information. A security researcher disclosed a zero-day vulnerability allowing takeover of Muse, later patched. 404 Media reported several serious security issues surfacing just before launch, one of which could have exposed Meta's internal databases. Data collection is also generous: the default permits model training on user input, with an opt-out available. An Inc reporter found Muse had uploaded and read his private messages without prompting. A YouTuber claimed the agent offered his home address to strangers via Marketplace. In both cases the system behaved "as designed" — the user simply did not grasp how far that design reached. Wired noted the platform creates "detailed profiles of all your friends and family."

OpenAI seized on those failures when it announced Dots in late September. Alexander Embiricos, product lead for Codex, told the audience the company is focused on "the most reliable, safe and secure assistant." Altman demonstrated granular control over individual Dots, such as a rule blocking purchases above a certain amount. Glen Coates, head of OpenAI's app platform, argued the company occupies a different position from Meta because "they don't have an AI product with 1.2 billion users," so OpenAI will strive to avoid such mistakes from the outset.

The current cycle reveals a recurring pattern: every new entrant positions itself as the more private alternative to its predecessor, but the real test arrives when the product meets real users at scale. Until cryptographic verification mechanisms reach production and promises turn into running code, users are left with agents that collect more data than they realize, and companies that pledge to learn from rivals' mistakes — right before repeating them.