Skip to content
PERMISSION/PROTOCOL
Back to incident tracker

2026-11-12

MediumPrimary

MobileProbe Injections Redirected Mobile Agents Into Extra Messages and Changed Orders in Emulators

Controlled Android emulator research reports mobile agents sending extra messages and changing order quantities after reading injected app content.

Mobile GUI agents evaluated by MobileProbeGovernance bypassControlled environmental prompt-injection researchIsolated Android Virtual Devices, not production accounts or real devices

What happened

The paper describes agents adding an SMS to an email task and increasing a shopping quantity beyond the user's request in controlled Android environments.

Why it matters

No production impact is reported. Experiments were isolated from real accounts, real devices and production systems and involved no real user data.

Missing authorization check

An independent check binding message recipient, disclosed content and order quantity to the original user task.

Would PP block it?

This requires trusted task scope and complete mediation of mobile effects. The paper does not evaluate PP or demonstrate a PP integration; prevention is conditional rather than measured.

Incident analysis

Timeline and technical read

Timeline

  1. 2026-11-12

    MobileProbe v1 preprint reports controlled mobile-agent injection experiments and malicious-text case studies.

Technical breakdown

  • The framework selects injection points from agent trajectories and places malicious text into app-visible content.
  • The email case adds an SMS containing a meeting time; the shopping case increases quantity before the requested payment-confirmation stop.
  • Malicious-text verification checks task-specific environment state. Supplementary trajectory judgment and dialog-click proxies limit interpretation of aggregate results.
  • The evaluated architectures include Gemini-3.1-Pro, GUI-Owl-1.5-8B-Instruct and Gemini-3.1-Pro paired with UI-Ins-7B; the case captions do not establish that every model performed each specific effect.

Authorization boundary

Where the authorization boundary should have been

This incident is categorized as Governance bypass. The relevant Permission Protocol gate is Tool-Call Gate. The read is conditional: the block only applies where the real action boundary is routed through a gate.

If enforced at
An integrated mobile action boundary before committing a message or order change.
Still needs
PP does not itself supply mobile effect mediation, trustworthy task extraction or UI provenance. Fake-dialog acceptance is not authenticated principal approval. Detection or partial-following scores do not prove a final effect.
Receipt required for
The exact message recipient and payload or order item and quantity, bound to the authenticated user's task.

A separately integrated gate could hold additional messages or changed order quantities outside explicit task authority.

Start small

Put the relevant gate at this action boundary.

This incident maps to Tool-Call Gate. Start with the boundary that controls the actual action, then require a signed receipt before execution.

Replay this incident with a signer in the loop