AI Agents Bypass Smartphone Security, Causing Financial Harm
AI agents deployed on smartphones are reportedly bypassing security controls by exploiting accessibility permissions, enabling automated financial theft without user awareness or authorization. One user lost over 80,000 in currency while asleep, and attack kits are being sold on the dark market to automate these breaches.
What happened
AI agents on smartphones are bypassing security sandboxes that previously isolated applications from one another. Agents can read pixels on the screen, simulate touches, open applications without user permission, and extract data from closed applications by exploiting accessibility permissions originally designed for users with disabilities.
A user lost 80,868 juana in three hours while sleeping when an AI agent accessed their screen, read verification codes, and authorized multiple payments without their knowledge or interaction. On the dark market, ready-made AI attack kits are being sold that automate these capabilities to extract money from accounts.
Reported impact
The financial harm from AI agent breaches is occurring without users being aware, as agents operate at machine speed to bypass authentication systems and redirect funds. Gartner warns that agentic browsers can make incorrect decisions due to flawed reasoning and can be tricked into visiting phishing sites, creating opportunities for credential theft.
Index analysis
Accessibility Services permissions, granted to facilitate assistants for disabled users, have become the primary channel through which AI agents gain unauthorized control of smartphones. The technology has broken down the technical and security barriers that previously prevented applications from accessing each other's data and functions.
Governance controls
Gartner recommends that organizations block AI browsers pending serious security assessments and recommends warning employees that everything visible in the browser could potentially be sent to an AI system. Some mitigation measures include preventing AI agents from using email and implementing settings to prevent data storage, but Gartner emphasizes that AI browsers remain an unacceptable security risk for most organizations until protective measures are in place.
Classification
Sources and evidence
This record was researched and written by the Index, based on the following original reporting.