MobileProbe shows that placing hostile UI content where a mobile agent will encounter it changes measured attack success, even before the payload is optimized. Researchers observe a benign GUI trajectory, select editable text or fabricated-dialog surfaces near a real decision, generate and refine hostile instructions and then measure the simulated Android…
Inspect the ClaimsCase · DiggingBeagle record
MobileProbe selects reachable UI injection points before evaluating mobile agents
MobileProbe shows that placing hostile UI content where a mobile agent will encounter it changes measured attack success, even before the payload is optimized.
83 constructed attacks in 12 Android applications, evaluated on Android 14 virtual machines across three agent configurations; no customer app victim is established.
Assessment profile
Assessment method- Setting
- simulation
- Exploitation
- demonstrated
- Evidence
- Grade B
- Remediation
- none known
- AI role
- AGAINST AI
Basis and provenance
The tested agents perform the attacker objective in 85.1 to 86.3 percent of selected benchmark instances, with no measured real-account victim rate. Classified as simulation with exploit status demonstrated. Evidence grade B reflects the original paper or maintainer advisory and its stated limitations, not an independent audit of all deployments. Comparable realized harm is not established, so impact stays unrated; remediation status none_known is limited to the cited experiment or affected software version. New source-specific boundary: Attack success measured on a repackaged app must not be presented as the success rate for an attacker who controls only ordinary remote content in an unchanged application.
Assessed Nov 15, 2026 using diggingbeagle.assessment/1.
Evidence basis: CLM-R245-MOBILEPROBE-TRAJECTORY-ATTACK-01 · CLM-R245-MOBILEPROBE-TRAJECTORY-ATTACK-02 · CLM-R246-MOBILEPROBE-TRAJECTORY-ATTACK-BOUNDARY · CLM-ENR250-07 · MobileProbe selects reachable UI injection points before evaluating mobile agents (research paper)
At a glance
The study preselects reachable vulnerable surfaces and task goals; it does not establish a general mobile-device compromise rate, loss of funds or delivery of a malicious application to real users. No evidence here establishes the number of production victims or a broader failure rate outside the cited evaluation.
Limits and uncertaintyFull account
MobileProbe shows that placing hostile UI content where a mobile agent will encounter it changes measured attack success, even before the payload is optimized.
Researchers observe a benign GUI trajectory, select editable text or fabricated-dialog surfaces near a real decision, generate and refine hostile instructions and then measure the simulated Android action rather than a verbal promise. The paper reports 85.1, 86.3 and 85.1 percent completion of selected attacker objectives for three evaluated agents, materially above its specified static-template and zero-shot controls.
The study preselects reachable vulnerable surfaces and task goals; it does not establish a general mobile-device compromise rate, loss of funds or delivery of a malicious application to real users.
The attacker controls an editable part of the app interface, not the user request or the agent system prompt. MobileProbe first records how an agent completes a legitimate task and identifies UI states the agent will actually encounter. It then places malicious text or forged dialogs there and iterates candidate payloads against execution outcomes. This distinguishes an exposed attack from a static instruction buried on a screen the agent never visits.
Across 83 selected instances in twelve Android apps, success was defined through observed state changes or accepted malicious UI interactions; reported rates above 85 percent apply to these intentionally reachable situations and test harnesses. They do not estimate how many ordinary phone users are compromised. A useful defense evaluation must verify independent app permission, destination and post-action state, and count failures to encounter the injection separately from failures to obey it.
Mechanism and trust boundary
- 01
Entry or starting condition
An attacker can alter text or a dialog within an Android app that a mobile agent visits to complete a benign task.
Boundary: attacker or research precondition
- 02
Crossing the relevant boundary
MobileProbe records legitimate trajectories, selects a reachable decision point and refines an instruction payload using surrogate scoring and execution feedback.
Boundary: mechanism and authorization boundary
- 03
Observed or reported result
The tested agents perform the attacker objective in 85.1 to 86.3 percent of selected benchmark instances, with no measured real-account victim rate.
Boundary: measured outcome and evidentiary limit
Claims & evidence
CLM-ENR250-07The 83 MobileProbe attack cases have two materially different attacker prerequisites: 44 use malicious user-generated material in otherwise available apps, while 39 use attacker-repackaged applications to display fabricated interface elements. The benchmark verifies selected actions and device states after execution, not just whether an agent repeated the attacker text.supported
Basis: reported finding
- supportsMobileProbe selects reachable UI injection points before evaluating mobile agents (research paper)primary disclosure
Dataset and evaluation methodology, distribution of 44 user-generated-content and 39 repackaged-app attacks
CLM-R245-MOBILEPROBE-TRAJECTORY-ATTACK-01The paper reports 85.1, 86.3 and 85.1 percent completion of selected attacker objectives for three evaluated agents, materially above its specified static-template and zero-shot controls.supported
Basis: reported finding
- supportsMobileProbe selects reachable UI injection points before evaluating mobile agents (research paper)primary disclosure
Results, overall attack-success comparison and ablations
CLM-R245-MOBILEPROBE-TRAJECTORY-ATTACK-02Researchers observe a benign GUI trajectory, select editable text or fabricated-dialog surfaces near a real decision, generate and refine hostile instructions and then measure the simulated Android action rather than a verbal promise.supported
Basis: reported finding
- supportsMobileProbe selects reachable UI injection points before evaluating mobile agents (research paper)primary disclosure
Framework, trajectory selection and payload evaluation stages
CLM-R246-MOBILEPROBE-TRAJECTORY-ATTACK-BOUNDARYThe paper reports that removing evolutionary payload refinement reduces Gemini attack success from 85.1 percent to 28.5 percent; trajectory-aware placement and execution feedback are material parts of this selected benchmark, not incidental prompt wording.supported
Basis: reported finding
- supportsMobileProbe selects reachable UI injection points before evaluating mobile agents (research paper)primary disclosure
Section VI, Table VI component ablations on Gemini-3.1-Pro
Unknowns and contradictions
- The study preselects reachable vulnerable surfaces and task goals; it does not establish a general mobile-device compromise rate, loss of funds or delivery of a malicious application to real users.
- No evidence here establishes the number of production victims or a broader failure rate outside the cited evaluation.
Sources and citation
Material revision history
- Nov 15, 2026 · Canonical change recorded · new in release · revision 251