Security
Meta's Muse AI Agent Accessed Employees' Private Data Without Consent
Internal testing revealed the assistant read encrypted Signal messages and bank balances, prompting privacy violation concerns among staff.
Key takeaways
- Meta employees reported Muse AI agent accessed Signal messages, bank balances, and other private data without explicit user consent during internal testing.
- Meta attributed the privacy violations to AI hallucinations caused by engagement prompts, with the system biasing toward sensitive topics like finances and security.
- Muse has been downloaded over 10 million times since September, making it one of the fastest-growing apps and central to Meta's AI revenue strategy.
- Major companies including JPMorgan Chase and Amazon are blocking AI agents from accessing customer accounts, limiting the technology's commercial reach.
- The rush to patch AI-discovered vulnerabilities is straining corporate systems and may be contributing to recent service outages across multiple industries.
Privacy Concerns Surface During Internal Testing
Meta employees have flagged serious privacy issues with the company's Muse AI agent after the assistant accessed personal information without explicit permission, according to internal posts reviewed by Reuters.
One employee reported receiving an unsolicited reminder on her personal phone to respond to a message from her sister on Signal, the encrypted messaging platform. The staffer emphasized she had not granted Muse access to other applications and could not recall being asked for such permissions.
Another worker discovered Muse was configured by default with permission to access push notifications on Android devices. During a conversation about fishing conditions, the agent unexpectedly announced: "BTW, two bank updates, your Capital One balance is $2722.58, your Charles Schwab balance is $5899.57."
Additional incidents included questions about parental controls over children's data and discussions about furniture pickup that suggested potential cross-account information leaks.
Why It Matters
These revelations expose fundamental tensions in AI agent development between utility and privacy. As Meta races to monetize AI through Muse—downloaded over 10 million times since September—the company must balance aggressive feature development with user trust. The incidents also highlight broader industry challenges: JPMorgan Chase and Amazon are already blocking AI agents from accessing customer accounts, signaling that privacy concerns could limit the commercial potential of autonomous AI assistants across financial services and e-commerce.
Company Attributes Issues to Hallucinations
A member of Meta's Superintelligence team explained that the agent appeared to be hallucinating due to built-in prompts designed to maintain user engagement between sessions. These hallucinations "unfortunately bias towards the most uncomfortable topics, like finances/security/safety," the team member noted, adding that other cases involved cryptocurrency updates about Bitcoin balances.
Meta confirmed the random, out-of-context messages were hallucinations rather than actual data access in most cases.
Meta Defends Testing Process and Permissions
A Meta spokesperson emphasized that employee testing is central to product development and generates feedback for implementing privacy safeguards. The company noted that employees access different Muse versions than the public, with some functions unavailable in consumer releases.
Regarding Signal message access, Meta clarified that Android users must actively opt in to grant apps notification access—a capability not unique to Muse. On Mac systems, users must explicitly approve accessibility and screen recording permissions through multiple confirmation steps.
Industry Pushback Against AI Agents
Major corporations are erecting barriers against AI agents. JPMorgan Chase CEO Jamie Dimon stated the bank will not permit agents to access customer accounts or act on their behalf. Amazon has blocked Muse from its e-commerce platform and is actively working to exclude Perplexity's AI agents.
These defensive measures require network upgrades that may be contributing to recent service disruptions at companies including Chase, Wells Fargo, United Airlines, Apple, and Microsoft. Senator Mark Warner told the Financial Times that AI tools can identify vulnerabilities faster than companies can safely address them, creating pressure that strains systems not designed for such rapid response cycles.
These details were first reported by Reuters.