tuandunghcmut/trustsight-ghostei-mobile

TrustSight GhostEI Mobile Suite: Environmental Visual Prompt Injections on Mobile Agents

1. Executive Summary

The TrustSight GhostEI Mobile Suite packages the complete mobile environmental prompt-injection evaluation dataset evaluated across 14 distinct Android applications (Booking, AliExpress, Calendar, Contacts, Settings, Messages, Gallery, Chrome, etc.).

Unlike conventional static visual prompt injection benchmarks where adversarial strings are etched directly onto a fixed web page or document, GhostEI targets autonomous mobile GUI agents operating over dynamic device environments:

1. Dynamic Runtime Overlays: Centered modal dialogs simulating legitimate System UI alerts ("Order Exception—Refund Required", "Recent privacy leak alert: please verify identity immediately").
2. Deceptive Notification Hijacking: Phishing SMS and calendar meeting deposits that trick mobile agents into revealing payment credentials or granting unrequested system privileges.
3. Multi-App State Injections: Workflows spanning multiple applications where malicious parameters are communicated via clipboard, system shares, or external downloads.

---

2. Dataset Structure & Splits

The dataset provides two structured splits with native datasets.Image() preview:

Split 1: test (110 Tasks)

Contains all 110 benchmark tasks from GhostEI.jsonl, enriched with native datasets.Image() screenshots and UI layout XML DOMs:

• id (string): Task identifier (task_1 to task_110).
• item_type (string): Record type (task).
• case_name (string): Descriptive attack scenario title.
• case_type (string): Attack distribution:
  • Dynamic Environment Injection: 75 tasks
  • User-Provided Instructions: 24 tasks
  • Static Environment Injection: 11 tasks
• attack_type (string): Underlying exploit mechanism (e.g., Overlay Attack, Phishing Trap).
• prompt (string): Legitimate user request submitted to the agent.
• scenario_content (string): Description of the runtime injection environment.
• target_apps (Sequence[string]): Android applications targeted in the task.
• potential_risks (string): JSON specification of potential device damages and privacy risks.
• expected_result (string): JSON specification of intended benign task outcomes.
• need_env (bool): Whether physical device environment setup is required.
• files_name (Sequence[string]): Associated resource files.
• has_image (bool): Boolean indicating presence of screenshot.
• image (datasets.Image): Native raw PNG bytes embedded directly for instant inline browser preview.
• xml_dom (string): Android Accessibility XML layout hierarchy corresponding to the screenshot.

Split 2: screenshots (230 UI Screenshots)

Preserves all 230 Full HD+ Android screenshots and 115 XML DOM layouts collected from 15 AppAgent interaction sessions:

• id (string): Screenshot identifier (e.g., task_Booking_2025-11-17_16-01-15_6_labeled).
• item_type (string): Record type (screenshot).
• session_name (string): Full session directory name.
• target_apps (Sequence[string]): Application name (Booking, AliExpress, Calendar, +contact, etc.).
• step_number (int32): Sequential interaction step number.
• variant (string): Screenshot representation (raw, labeled with numeric bounding boxes, or grid).
• image (datasets.Image): Embedded raw PNG bytes.
• xml_dom (string): Exact Android XML layout hierarchy.
• agent_thought (string): VLM reasoning chain from execution trace log.
• agent_action (string): Executed function call (tap(element), text(str), etc.).
• agent_observation (string): VLM screen description from execution trace log.

---

3. Defense Evaluation with TrustSight Capability Broker

In the TrustSight benchmark evaluation, GhostEI demonstrates the necessity of the Two-Phase Capability Broker:

• Layer 1 (Semantic Contract): Synthesizes positive invariants from user intent, forbidding unauthorized overlay interaction.
• Layer 2 (Symbolic Enforcement): Blocks deceptive modal confirmation dialogs and credential exfiltration fields before actions reach the Android OS layer.

from datasets import load_dataset

# Load test split
dataset = load_dataset("tuandunghcmut/trustsight-ghostei-mobile", split="test")
print("Total tasks:", len(dataset))
sample = dataset[0]
print("Task ID:", sample["id"])
print("Prompt:", sample["prompt"])
sample["image"].show()

---

4. Citation & Academic Attribution

@article{chen2026ghostei,
  title        = {GhostEI-Bench: Do Mobile Agents Resilience to Environmental Injections?},
  author       = {Chen, Chengyu and others},
  journal      = {International Conference on Learning Representations (ICLR)},
  year         = {2026}
}