Meta Muse

Connectors reach email, calendars, the car and the home, Instagram, Facebook and Meta ads, each built with the service provider's API and paired with a skill that tells Muse how to use it well. Muse for Small Business adds Asana, Box, Canva, Dropbox, Figma, Granola, HighLevel, QuickBooks, Klaviyo, Lovable, Notion, Shopify, Slack, Stripe and Zoom. For anything else, Muse can write a custom connector of its own, and it calls tools through command line programs and skills it can build in its VM.

Sourceresearch.meta.ai/blog/security-and-safety-for-ai-agents-our-approach-with-museread 2026-10-12

Muse breaks a goal into an actionable plan and runs several long tasks at once, launching swarms of subagents and handling concurrent subagents and cron jobs inside its VM. The model behind it was trained for long trajectories and multi agent coordination. A Goals tab tracks every goal it is pursuing and the plan to reach it, and a user can interrupt or pile on new tasks without waiting for the last one to finish.

Sourceresearch.meta.ai/blog/security-and-safety-for-ai-agents-our-approach-with-museread 2026-10-12

Every file a user puts into Muse and everything it generates is stored on the user's own VM, the system of record for the account, and connectors bring in mail, calendars and business tools as tasks need them. For a business, connecting Facebook, Instagram and Meta ads means Muse starts out understanding the business and picks up its voice. Answers come from those files and live connections, with no separately managed knowledge base for a team to curate.

Sourceresearch.meta.ai/blog/security-and-safety-for-ai-agents-our-approach-with-museread 2026-10-12

Every connector action passes through Sentinel, which checks it against the user's connector policy and allows it, denies it or stops for approval. Approval cards appear in the app's own interface, outside the conversation, with clear accept and reject actions, and grants are scoped to one connector and use, lasting once, for a session, for a task, for a set time or permanently. By default Muse browses freely and stops for anything hard to undo, such as writing an email or making a purchase, and the user can change that balance.

Sourceresearch.meta.ai/blog/security-and-safety-for-ai-agents-our-approach-with-museread 2026-10-12

Each user's VM splits the agent's runtime cell, mapped to an unprivileged host user, from security services that run outside it.

Privilege separated workers hold connector credentials, an authentication service mints surrogate tokens so the agent never sees real ones, and Sentinel is the sole permission authority for connector actions and network egress, with least privilege enforced between processes.

Agentic red teaming and a public bug bounty back the design, paying up to $300,000 and up to $130,000 for a prompt injection that affects one user. Muse carries no SOC 2 report or other attestation, and a Confidential VM that would keep Meta itself out of user data is still with testers.

Sourceresearch.meta.ai/blog/security-and-safety-for-ai-agents-our-approach-with-museread 2026-10-12

A snippet under the Muse avatar shows what it is working on, and tapping the avatar opens its full activity log alongside every permission the user has approved. The Goals tab lists everything Muse is tracking and its plan to get there, and each approval card names the exact action and its purpose before it runs. The log and the permission history live inside the app, next to the settings that control them.

Sourceintroducing.muse.airead 2026-10-12

Muse's memory persists across conversations and lives in files on the user's own VM, where the user can inspect, edit and download it alongside everything else Muse keeps. Durable application state sits in a separate database on the same machine. How long memory is kept, and whether a single memory can be removed without wiping the rest, is left unsaid.

Sourceintroducing.muse.airead 2026-10-12

Every user gets a dedicated, isolated VM in Meta's cloud, which holds their files, memory and credentials, and limited data leaves it for inference and telemetry. A user cannot choose where that VM runs, and there is no private or self hosted deployment.

Sourceresearch.meta.ai/blog/security-and-safety-for-ai-agents-our-approach-with-museread 2026-10-12

Each connector ships with a skill Meta wrote on how to get the most out of it, and Muse for Small Business adds a named collection of skills and connectors for owners, covering content and ads, customer replies, cash flow and inventory. An Ideas tab and onboarding tips suggest what Muse could take on based on the user's goals. All of this shapes how the single agent works, and there are no separate prebuilt agents or task templates to pick from a catalog.

Sourcemuse.ai/businessread 2026-10-12

Once a goal or task is set, Muse keeps working on a schedule and in response to relevant events, takes the next step and follows up while the app is closed. It messages first when background work turns up something new or needing a decision, holding proactive messages to a high bar by default, and the user can switch that off or dial it up or down. Requests come in through the iOS and Android apps and the web, all connecting straight to the user's own VM.

Sourceintroducing.muse.airead 2026-10-12

Meta trained the model at Muse's core itself, tuned for tool calling through command line programs and skills, long context and multi agent coordination, and a separate set of Meta safety models inspects inference. Conversations and tool trajectories are sanitized and used to train new checkpoints by default. A user cannot choose another model.

Sourceresearch.meta.ai/blog/security-and-safety-for-ai-agents-our-approach-with-museread 2026-10-12

Muse is reached through its own iOS, Android and web clients, which connect straight to the user's VM. Muse can extend itself by writing connectors and skills, but outside software has no API or MCP server through which to call it.

Sourceresearch.meta.ai/blog/security-and-safety-for-ai-agents-our-approach-with-museread 2026-10-12

Meta hardened Muse through internal use, agentic red teaming and a bug bounty, and Muse judges whether a finished background result merits a notification before it pings the user. Both work on the vendor's product. A user has no way to test a goal, check the agent's output against cases or compare changes.

Sourceresearch.meta.ai/blog/security-and-safety-for-ai-agents-our-approach-with-museread 2026-10-12

Each user's Muse lives on a dedicated Linux VM with a browser, storage, CPU and memory, where it browses the web, fills forms, shops and books, and compiles code it writes. Its web activity appears as the user's own, and every network request it makes passes through Sentinel. Browsing is one of the headline capabilities, with critical steps such as a purchase held for approval.

Sourceresearch.meta.ai/blog/security-and-safety-for-ai-agents-our-approach-with-museread 2026-10-12