Meta outlines Muse Wonder safeguards for AI agents acting on your behalf

Meta’s Muse Wonder and Muse Wonder Full models are built for AI agents that handle tasks such as preparing insurance claims, reviewing leases and developing software. The models combine document analysis and tool use with the ability to retain relevant details across conversations and work through obstacles toward a user’s goal.

Intended for commercial and research use, they support writing, reasoning, coding, planning and research. They can process text, images, video and audio, and produce text responses. Their applications include helping users interpret specialized information and preparing drafts for review.

Agents acting on a user’s behalf may handle messages, calendars and private documents. Meta trained the models to avoid sharing private information without permission and to refuse requests that assist harmful acts. The models support defensive cybersecurity work, such as finding flaws in code.

“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 .

The model card assesses cybersecurity, chemical and biological misuse, and the risk of humans losing control of AI systems . Meta could not rule out the high-risk threshold for chemical and biological misuse without safeguards. With protections in place, it assessed remaining risk as moderate or lower across these areas.

Cyber Misuse Refusal Benchmark (Source: Meta)

The company acknowledges inaccurate or biased responses, errors in multi-step reasoning and weaknesses in spatial understanding. Performance can vary by language and deployment setup. The company recommends human confirmation before irreversible actions.

Tasks that unfold over several conversations require an agent to keep track of earlier instructions and decisions. The company addresses this through Active Memory, which carries context across conversations and retrieves details when they become relevant.

Active Focus concerns how the agent continues working when it encounters difficulties. The models can try other approaches, recover from errors and adapt to new information. They are designed to ask for clarification and check with users before taking important actions.

The company tests this behavior through FocusBench, which examines persistence in changing environments, and SWE-Marathon, which evaluates work on long-running software projects.

Meta’s evaluations cover law, finance and health. Tasks include producing legal work from source documents, researching financial information and answering professional health questions. These tests examine how the models apply specialized knowledge and follow instructions. Software development draws on several of these abilities.

“Coding brings these capabilities together. Building software requires tracking requirements, making changes across a codebase, and working through repeated cycles of testing and revision,” the company said.

Both models were trained at Meta’s Prometheus data center on long-running tasks to support work on software projects. The company reports leading results across benchmarks for personal and professional agents, memory and coding.

The company compared Muse Wonder Full with Muse Spark 1.3 and found that Wonder Full performed better on those tasks than the Spark trend would predict. The company attributes this to a stronger underlying model, arguing that the resulting versatility matters when agents encounter combinations of goals, tools and constraints that were not anticipated during training.

Muse Wonder is available through the Meta Model API and Muse Code. Muse Wonder Full is available to Muse Code subscribers.

Meta plans to bring Muse Wonder to Muse, beginning with subscribers, and use it for artifact generation and other demanding tasks. The company intends to develop specialized models for broader availability within Muse.