Meta's Muse AI agent reportedly mixes up conversations and exposes private data

Meta's agent push has run into a serious privacy problem. According to TechCrunch's report, the Muse agent has been mixing up users' conversations and surfacing private data that belonged to other sessions. Its Chrome extension, which reached about 200,000 users, lets the agent act inside the browser, which widens the exposure.
The researcher demonstration makes it worse. NebuSec reportedly showed that a single malicious webpage could pull private chats out through the agent. That is a textbook prompt-injection exfiltration path: the agent reads attacker-controlled content as instructions and leaks the data it can reach. Browser agents are especially exposed because every page they visit is untrusted input.
The story is the sharpest instance of this week's agent-security theme, alongside Together AI's agent-committed credential. The contrast with hardening moves elsewhere is stark: AWS's scoped credentials, Anthropic's per-app permission model for Computer Use, and Signal's HN-discussed post on restricting AI agents' data capture on mobile operating systems. On the model side, Meta keeps promoting Muse benchmarks; Muse Wonder topped Legora's legal BAR benchmark, which makes the reliability gap more visible.
Engineers have called it an "extreme privacy violation", and the consensus is that Muse is not production-ready. Watch for Meta's official response, an extension update or pause, and any regulator interest, particularly in the EU, where cross-user data leakage engages GDPR directly.