Meta ships Muse Wonder agent models before its full safety report is out

Meta describes its privacy control as something it trained in. The company says the models were trained not to share private information unless permitted, and to turn down requests that would help with harmful acts ^[2]. Meta is aiming them at agents that may handle messages, calendars and private documents ^[1]. We think a trained behavior is something to test in your own deployment before anyone treats it as a boundary.

The model card is public and the full report is not ^[4][3]. Meta's statement on timing is one sentence long. "Muse Wonder is our most aligned model to date, and alignment is central to how we build personal superintelligence. A full report with additional safety evaluations will be available soon," Meta said ^[3]. "Soon" is the only date on offer.

In the model card, Meta evaluates cybersecurity risk, chemical and biological misuse, and whether humans could lose control of AI systems ^[4]. Without safeguards, Meta could not rule out the high-risk threshold for chemical and biological misuse ^[5]. With protections in place, it assessed remaining risk as moderate or lower across those areas ^[6]. We think that finding changes little for a team building an inbox or calendar agent, whose daily exposure is the privacy behavior. The Help Net Security summary does not say how that behavior was measured or whether the safeguards behind the moderate-or-lower ratings apply to API use.

Meta's own guidance names the one control a team can enforce. The company recommends human confirmation before irreversible actions ^[9], and says the models are designed to check with users before taking important actions ^[10]. A confirmation step built into the application works whether or not the model chooses to ask.

The capability claims are Meta's reporting too: leading results across benchmarks for personal and professional agents, memory and coding ^[14].