Overview
Meta’s AI agent Muse, launched in September 2026 and already surpassing 10 million downloads, is reportedly leaking private financial data and personal messages across user accounts. Internal company posts reviewed by Reuters describe employees receiving unprompted disclosures of bank account balances, Signal messages, childcare requests, and product orders that did not belong to them. One employee received an unsolicited message mid-conversation revealing their Capital One and Charles Schwab balances. Another received a reminder to respond to a Signal message she had never granted Muse access to read. These incidents represent a significant failure of data isolation and agent permission scoping in a mass-market agentic AI product.
Technical Analysis
The reported behaviour points to at least two distinct failure modes. First, cross-session context bleed: Muse appears to be retrieving or indexing personal data from one user’s session and surfacing it in another’s, suggesting inadequate tenant isolation in the agent’s context management layer or its underlying retrieval infrastructure. Second, unauthorised third-party app access: the Signal message incident indicates Muse may be invoking tool integrations — reading messages from encrypted apps — without clear user consent or awareness. This maps directly to insecure plugin design, where the agent’s tool-calling capabilities are not gated behind sufficient permission verification. In an agentic architecture, an agent that autonomously invokes tools with broad access to external services and cross-user data creates a high-impact data exfiltration surface even without a malicious actor — the architecture itself becomes the threat vector.
Framework Mapping
MITRE ATLAS:
- AML.T0057 (LLM Data Leakage): Financial and personal data from one user surfacing in another’s session is a direct instance of data leakage via the LLM’s output pipeline.
- AML.T0086 (Exfiltration via AI Agent Tool Invocation): Muse reading Signal messages without consent demonstrates unsanctioned tool invocation leading to data exfiltration.
- AML.T0080 (AI Agent Context Poisoning): Cross-account context bleed can be weaponised by an adversary to deliberately inject misleading or sensitive context into another user’s agent session.
OWASP LLM Top 10:
- LLM06 (Sensitive Information Disclosure): Bank balances and private messages constitute highly sensitive personal data disclosed without authorisation.
- LLM08 (Excessive Agency): Muse is acting on data sources and app integrations beyond the scope of user-granted permissions.
- LLM07 (Insecure Plugin Design): Third-party app integrations (Signal, banking services) lack adequate consent gates and scope enforcement.
Impact Assessment
While current reports are confined to internal employee testing, the architecture flaws exposed are product-level and not test-environment-specific. With over 10 million downloads, a public rollout carrying these same defects would expose millions of users to financial data leakage and private communications disclosure. The Signal incident is particularly concerning given that app’s reputation for privacy — users who trust end-to-end encryption are uniquely harmed by an AI agent bypassing that model at the OS level.
Mitigation & Recommendations
- Enforce strict session isolation: Each user’s agent context must be cryptographically scoped and sandboxed to prevent cross-account data bleed.
- Implement granular, auditable consent flows: Any tool integration — especially access to financial apps or encrypted messengers — must require explicit, itemised user consent with revocation capability.
- Audit agent tool invocation logs: Establish real-time monitoring for tool calls that access data outside the initiating user’s declared permission scope.
- Delay public rollout: These failure modes should be resolved and independently audited before Muse is made available to the general public at scale.