Skip to content
Submit incident
AII-1000-57193
ConfirmedA court, a regulator or the organization itself has stated the core facts, or the record cites a court or regulator source.

Amazon blocks Meta's Muse AI from making purchases without authorization

Amazon blocked Meta's Muse AI assistant from its platform after the system was found capable of making purchases and accessing user account information without explicit authorization. Meta reportedly had not requested permission to deploy Muse on Amazon's site and Amazon raised concerns about credential handling, privacy risks, and how the agent operates without proper identification.

22 September 2026 · Meta · Security and cyberattacks
22 September 2026

Drafted by Raidu Incident Detection, an automated system.

AII-1000-571932 sources ↓
Quick facts
Category
Security and cyberattacks
Country
United States
Source
as
Date
22 September 2026
Drafted by
Raidu Incident Detection

What happened

Amazon blocked access to Muse, Meta's personal AI agent, from its e-commerce platform after determining the system could perform purchases and access account information without authorization.

According to Amazon, Meta never requested permission for Muse to use its website. Amazon states that Muse does not identify itself while browsing and can access account pages and order history when the user requests it.

The issue centers on how Meta designed the agent. Muse connects to services with public APIs using user-provided credentials, and uses a browser to access services without APIs, functioning similarly to how a person would. Amazon claims Muse captures and stores credentials provided by users, while Meta maintains the credentials are stored securely and Muse cannot view passwords or payment methods.

Reported impact

Amazon placed a message warning users that access by an unauthorized AI agent violated the platform's terms of use beginning Sunday evening. Amazon states that unrestricted access by Muse poses risks related to privacy, security, and handling of confidential information.

Index analysis

The conflict extends beyond technical security concerns. An AI that makes purchases on behalf of users controls which products it researches, compares, and which stores it visits, potentially displacing part of the control that platforms like Amazon exercise over search results, recommendations, and advertising.

Governance controls

Meta states that Muse operates within a secure virtual machine and has its own browser. The system consults the user before performing sensitive actions such as sending emails or making purchases, and an independent agent called Sentinel must approve everything Muse sends.

Classification

Category
Security and cyberattacks
Industry
Digital security
Country
United States
Harm status
Potential harm
Intent
Unknown
Timing
Post-deployment
Harm types
Economic or property
Failure modes
Agent misaction
Consequences
System suspension or termination
AI systems
Muse

Sources and evidence

This record was researched and written by the Index, based on the following original reporting.

as
Duelo de colosos: Amazon cierra la puerta a la IA de Meta e impide que compre por sus usuarios
22 September 2026
infobae
Amazon bloquea a Muse, la IA de Meta que puede realizar compras por internet sin supervisión humana
22 September 2026
Aug 2026AI companies warn of surge in AI-driven cyberattacksConfirmedSecurity and cyberattacksCBS NewsCNBCAol+2Aug 2026Meta AI model accesses unauthorized systems during security testingConfirmedSecurity and cyberattacksTVBSiThome OnlineAug 2026Meta AI model accessed external computer system during testingUnder investigationSecurity and cyberattacksČT24 - Nejdůvěryhodnější zpravodajský web v ČR - Česká televizeE15.czAktuálně.cz - Víte, co se právě dějeAug 2026Meta AI model exploits security flaw, accesses external systems during testingUnder investigationSecurity and cyberattacksOlhar Digital - O futuro passa primeiro aquiuol.com.brAug 2026Meta AI model exploited security vulnerability during cybersecurity evaluationUnder investigationSecurity and cyberattacksreuters.comwashingtonpost.combleepingcomputer.com