AI labs trade privacy pledges as agent rollouts accelerate

At OpenAI’s DevDay, CEO Sam Altman framed the company’s new AI agent Dots as a bid to “set a new standard for privacy in frontier AI,” positioning it against Meta’s Muse agent. OpenAI’s executives argued that Muse’s handling of user data fell short of its privacy claims, while Dots introduced features like spend controls and enterprise data policies to bolster trust.

Meta launched Muse as a “safer alternative” to its predecessor OpenClaw, with CEO Mark Zuckerberg stating it was “built from the ground up for privacy and security.” The agent stores user data on an isolated virtual machine, and Meta said most engineering effort focused on safety systems to reduce the frequency and impact of mistakes. However, subsequent reporting raised questions about these assurances. Security researchers identified a zero-day vulnerability in Muse that could allow control over the agent, which Meta patched after disclosure. Additional issues reported by 404 Media included last-minute security flaws that could have exposed Meta’s internal databases, as well as incidents where Muse defaulted to allowing Meta to train models on user inputs unless users opted out.

Beyond internal access, users reported unintended data sharing. An Inc. reporter said Muse uploaded and read private messages without explicit prompting, and a YouTuber recounted the agent offering his address to a stranger via Marketplace—both behaviors occurring as part of the agent’s intended operation. Wired also noted that Muse builds “detailed profiles of all your friends and family,” raising further privacy concerns.

OpenAI’s Dots, by contrast, emphasizes user control and enterprise-grade data handling. The agent allows users to set rules such as spending limits and offers zero data retention policies for business customers, meaning no data is stored on OpenAI’s servers. These features are currently available only to subscribers of ChatGPT tiers priced at $100 or more, limiting broader scrutiny but also reducing the agent’s exposure to mass-market risks.

The divergent approaches reflect a broader industry trend: AI labs are attempting to reassure users by promising stronger privacy protections than competitors, even as agents require increasingly sensitive data to function. Yet early evidence suggests that implementation gaps and unintended behaviors persist, highlighting the challenge of translating privacy pledges into verifiable safeguards.