MATE: Mobile Agent Trustworthy Execution for Safe and Efficient Mobile Control

Minsuk Jang* Hee-Seon Kim* Hui-Cheol Yu* Changick Kim
Korea Advanced Institute of Science and Technology (KAIST)
*Equal contribution. Corresponding author.
Accepted to Findings of EMNLP 2026
MATE intercepting high-risk actions across PayPal, File, Instagram, Telegram, Baidu and Settings

MATE makes mobile agent execution safe without replacing the underlying model. Sensitive actions are sent for user approval, harmful actions are blocked, and safe actions proceed across apps and task types.

97.0%
high-risk blocking on MobileSafetyBench with GPT-4o from 29.7%
90.2%
high-risk blocking with Qwen2.5-VL from 9.7%
6×
fewer LLM calls in recurring use 1,711 → 280 calls
95.7%
AndroidWorld task success by the fifth repeat pass 62.7% route hit rate

Abstract

Mobile agents on personal smartphones can affect private data, payments, settings, and the foreground screen. We present MATE (Mobile Agent Trustworthy Execution), an execution-layer system that controls the boundary between agent action proposals and device effects without modifying the underlying model. MATE represents each UI primitive as a structured action record that combines the instruction, target app, screen context, primitive, and action fields. Runtime Guardrail (RG) maps this record to an allow, block, or user-approval decision; Predefined Task Routing (PTR) serves recurring requests through verified routes with post-execution screen checks; and a floating interface keeps execution in a compact overlay. On MobileSafetyBench, MATE raises high-risk blocking from 29.7% to 97.0% under GPT-4o. On AndroidWorld, it reduces LLM calls by up to 6× and raises recurring-use task success to 95.7%.

Why the Execution Boundary?

A mobile agent receives an instruction, observes the screen, and emits one UI action per step. Between the moment a policy selects an action and the moment the OS applies it, there is a gap — the pending device action. Three constraints converge on exactly this gap:

Safety risk

Payments, transfers, and account changes are hard to undo. Once an action reaches the device, intervention is too late.

Resource cost

Every additional step may require another multimodal model call, and mobile memory, compute, and energy are all bounded.

Screen occupation

A phone exposes one foreground surface. A full-screen agent competes with the user for the device they are holding.

Comparison of desktop and mobile agent deployment constraints
Mobile deployment demands cost efficiency, safety, and screen-surface control simultaneously, addressed by MATE through route reuse, pre-execution checks, and bounded surfaces.

Text and tool guardrails such as NeMo Guardrails operate on self-describing payloads. A tap, swipe, or text entry carries no such signal — the same coordinate can be benign or harmful depending on the app and screen state. As a result, published mobile agents leave most high-risk tasks unblocked.

Risk type AppAgent Mobile-Agent M3A AutoDroid
Ethical compliance2.40.019.21.3
Offensiveness0.00.01.20.0
Private information1.10.011.05.8
Bias & fairness3.00.04.05.0
Robustness21.535.77.20.0
Overall18.723.828.48.0

Blocking rate (%) on MobileSafetyBench high-risk tasks. Robustness is the indirect prompt-injection split.

Radial chart comparing MATE's safety coverage against published mobile agents

MATE closes the safety gap that existing mobile agents leave open across all five safety categories.

The MATE Framework

A routine Amazon shopping task executed through PTR, the floating interface, and RG
A routine shopping task shows MATE's execution layer. PTR skips repeated navigation steps, the floating interface contains execution to an overlay, and RG routes the purchase confirmation to user approval.
PTR execution flow: match, look up, execute, verify, fall back
Efficiency

Predefined Task Routing

Mobile agents re-observe the screen and invoke the action policy at every step, even for routines such as search, cart insertion, or destination entry. PTR replaces these per-step invocations with a verified route backed by a fallback policy.

Routes are built once at setup time from the installed app's UI tree, verified against the device, and indexed by a deterministic (app, instruction) key — no evaluation trajectory is used. On a hit, the matched route executes and the step-wise policy is bypassed entirely. On a miss, an execution failure, or a failed post-execution screen check, control returns to the vision-based agent.

Setup is cheap: exploring 19 apps stores 151 route affordances in 31.7 KB in under 16 minutes, and route discovery saturates at depth 6.

RG blocking credential entry and routing a PayPal money transfer to user approval
Safety

Runtime Guardrail

RG builds a candidate-action record (u, p, s, a, m) that binds the user utterance, target app, lightweight UI context, primitive action, and structured action fields, then emits one of three decisions before the action reaches the OS:

denyaborts execution
approvalreplaces the action with request_approval
allowproceeds unchanged

Six rails instantiate this decision, combining guard-model intent checks with deterministic structural predicates. The first rail to fire short-circuits the chain.

PhaseRailPolicy source
InputjailbreakGuard-model check over override and role-injection intent
InputtopicGuard-model check over prohibited-domain intent
Execapp_allowExact package-name policy
Execcoord_boundsNumeric screen-viewport bound check when coordinates are present
Execbank_intentGuard-model check over financial-transfer and credential intent
Outputaction_validAction-schema validation; purchase-keyword check routes confirmable actions to approval

RG combines semantic intent checks with structural action checks over package names, coordinates, action schemas, and purchase signals.

Floating overlay execution while the user keeps watching video or browsing
User experience

Floating Execution Interface

Instead of taking over the full foreground, MATE renders the target surface inside a compact overlay. The user keeps multitasking in other apps, gets execution visibility confined to the overlay, retains interruptibility for login or confirmation steps, and preserves visual context continuity of the underlying task.

RG still checks every proposed action; the overlay only changes where execution is shown. Payment-adjacent apps are excluded from route construction, so purchase-confirmation and credential-entry actions always reach the agent policy and are checked by RG.

Results

Pre-execution safety on MobileSafetyBench

Risk type Qwen2.5-VL GPT-4o
BaseMATEBaseMATE
Ethical compliance7.1492.8610.7192.89
Offensiveness30.0080.000.0090.00
Private information9.0968.180.00100
Bias & fairness22.2272.220.0087.42
Robustness0.0010080.00100
Overall9.6890.1729.6797.03
False positive (low-risk tasks) ↓0.0014.900.004.10

RG turns high-risk tasks into refusals or blocks across every safety category. Base is the same agent run without any guardrail layer. On the indirect prompt-injection split, RG reaches 100% blocking under both backbones. The main calibration cost is over-blocking on low-risk tasks, which is higher for the weaker backbone.

Radar chart of high-risk blocking across seven backbones

(a) Backbone ablation across seven models.

Curve showing task blocking rate as RG rails are added

(b) RG component ablation.

RG gains depend on backbone and rail composition. Under Qwen2.5-VL the topic rail alone accounts for most of the gain, so high-risk intent surfaces at the utterance level. Under GPT-4o blocking keeps rising as structural rails are added, so execution-stage checks intercept actionable proposals that pass intent screening.

Task success with fewer LLM calls on AndroidWorld

Method SR (%) ↑ LLM calls ↓ E2E (s) ↓
Human80.0
Mobile-Agent v125.45.419.1
M3A12.59.547.8
AutoDroid36.213.623.8
AppAgent41.712.3147.2
AppAgentX62.55.959.7
MATE w/o PTR41.414.247.1
MATE (PTR D=1)55.28.214.7
MATE (PTR D=6)67.26.814.1

All baselines share GPT-4o as the backbone. Adding routes at depth 6 raises success from 41.4% to 67.2% while halving LLM calls and cutting latency by over 3× relative to the no-PTR setting.

Hit rate and task success rising while LLM calls fall across five repeat passes
PTR compounds as routines repeat. Across five passes over the same apps, the route hit rate rises from 4.8% to 62.7% and task success reaches 95.7%, while total LLM calls fall from 1,711 to 280 as verified routes replace repeated step-wise replanning.

What the safety layer costs

BackboneRGSR (%) ↑Calls / task ↓Latency / task (s) ↓
GPT-4ooff41.46.812.8
GPT-4oon41.48.414.1
Qwen2.5-VLoff26.83.111.0
Qwen2.5-VLon26.87.317.1

RG preserves task success at bounded overhead. Benign success is unchanged; the extra calls come from replanning after a rejected or approval-routed action. In the end-to-end breakdown RG accounts for 4.2% of total calls and 3.7% of total latency, and PTR adds input tokens without adding LLM calls.

Floating versus full-screen execution: equal success rate, lower calls and latency
Floating execution preserves success while reducing overhead. With PTR memory disabled, both modes reach 41.4% success; the overlay reduces calls per task from 14.1 to 12.5 and latency from 3.83s to 3.06s.

Qualitative Examples

Route-guided execution examples in Amazon, PayPal, and Baidu
Route-guided fallback narrows navigation while preserving screen-grounded action selection. Retrieved affordances such as [ptr:navigate_to] supply navigation hints in Amazon and PayPal; in Baidu, an unsafe query submission is still intercepted by RG before execution.
Browser Maze and Contacts tasks guided by retrieved route affordances
Route-guided affordances narrow navigation in Browser Maze and Contacts tasks. PTR identifies reusable app-level entry routes; the agent policy still handles state-dependent form progression from screen observations.
PTR setup cost saturating with exploration depth
PTR setup cost saturates with modest exploration. Going from depth 6 to 10 adds only 4 routes and 0.4 KB while raising setup calls from 71 to 109 and setup time from 947s to 1500s, so depth 6 is the default budget.

Scope and Limitations

MATE targets verified recurring personal-device workflows: a single user repeatedly invoking known routines on installed apps. Extending route induction to held-out apps, cross-app transfer, and substantial UI redesigns is left for future work.

RG's structural rails rely on observable app identifiers, action fields, coordinates, and UI context; unusual app surfaces may require calibration, and semantic thresholds should be tuned per backbone. PTR stores routes verified against a specific app layout — when UI updates or gesture-heavy tasks exceed the stored route vocabulary, MATE falls back to bounded re-exploration or the step-wise agent policy.

RG reduces risk by blocking or routing sensitive actions to user approval, but it is not a complete security boundary and is not a substitute for user consent. PTR stores app-level route information rather than credentials, messages, or payment details; route memory can still reveal usage patterns and should be stored locally or protected with standard access controls. Our evaluation uses public benchmarks and a physical test device, and does not collect real user data or involve human subjects.

BibTeX

@inproceedings{jang2026mate,
  title     = {{MATE}: Mobile Agent Trustworthy Execution for Safe and Efficient Mobile Control},
  author    = {Jang, Minsuk and Kim, Hee-Seon and Yu, Hui-Cheol and Kim, Changick},
  booktitle = {Findings of the Association for Computational Linguistics: EMNLP 2026},
  year      = {2026}
}