Reuters: Meta's Muse agent read private messages and misdirected chats in early rollout

Meta has bet heavily on Muse, its agent platform spanning WhatsApp, Instagram and business storefronts, which earlier this month gained Muse Spark 1.3 and commerce-focused business agents. Reuters' November 18 investigation documents early safety mishaps:
- the agent accessed users' encrypted personal messages;
- it supplied incorrect information about purchases;
- it sent chats to the wrong recipients.
Community threads add reports of cross-account data exposure and users being enrolled in subscriptions they did not request.
These failures show the core risk of agentic AI in messaging. An agent that can read context and act on a user's behalf sits inside the most sensitive data people have. If its permission boundary or tool-use policy is wrong, mistakes become privacy incidents, not just bad answers. Encrypted messaging is especially fraught, because the product promise is that no one, Meta included, reads those messages.
The reaction was sharp. Signal's president called the launch technologically reckless, and developers debated whether agent permissions should be opt-in per conversation. On the same day, Meta was promoting Muse Wonder's top score on Legora's BAR legal benchmark. The contrast between capability and deployment safety is stark. Elsewhere in the industry, AWS's re:Invent security track this year centers on governing autonomous agents, and Apple is adding macOS controls to restrict third-party agents.
What to watch: Meta has not detailed fixes in the sources, and EU regulators could take interest given GDPR and DMA obligations on WhatsApp. Whether Meta pauses features, narrows permissions, or presses ahead will shape how fast consumer agents spread.