On September 8 (local time), Meta announced the personal AI agent "Muse." Muse is equipped with the ability to not only answer questions but also execute actual tasks on behalf of the user. Primary use cases include sending emails, booking travel, and opening web browsers to fill out forms; the agent can autonomously create and execute plans based on a given goal.
Meta Announces Personal AI Agent "Muse" Featuring Task Automation and Advanced Security
This article is a translation. Read the Japanese original
Its foundation model is "Muse Spark 1.3," which the company describes as its highest-performing model to date. In addition to a dedicated app, Muse supports instructions via WhatsApp and is now available on iOS, Android, and the web. A subscription plan is also expected to be available for advanced users.
Regarding security, each user is assigned a dedicated virtual machine in the cloud (Muse Secure VM), ensuring that data and authentication information are stored within an environment isolated at the system level. Furthermore, a separate agent called "Sentinel" resides in the system to monitor and authorize operations and network communications with external services, a design intended to prevent unintended actions. For payments, Muse can utilize Stripe's digital wallet "Link," which issues single-use card numbers to limit potential damages in the event of an incident.
Meanwhile, Meta acknowledged the risk of malfunctions caused by prompt injection (misuse of instructions) as an industry-wide challenge. In response, Meta has implemented multi-layered defenses combining input data filtering and human approval processes. On the same day, the company also opened its bug bounty program to the public, stating it will pay rewards of up to $130,000 for successful examples of prompt injection.
Sources
- Meta、パーソナルAIエージェント「Muse」発表 メール送信や買い物も代行（まずは米国で無料提供開始） （ITmedia AI+, 2026-09-09）
- Meta 公式ブログ