App Threat Report Q2 2025: Financial app security in 2025: Combating traditional malware and emerging AI threats PDF Free Download
1 / 43
/43
100%
Prepared by: Promon Security Research Team
Combating traditional malware and
emerging AI threats
App Threat Report
Q2 2025
06 Europe
09 The United States of America
10 Southeast Asia
12 BingoMod
13 ToxicPanda
13 Android malware
17 Deepfake threats in biometrics
18 Why AI only solutions fall short in mobile deepfake video detection
19 The case for layered defense
20 Conclusion: Beyond AI alone
22 Financial impact shows urgent need for protection
23 The AI security challenge
23 Critical AI threats on devices
25 How can your AI applications be protected?
26 Regulatory compliance benefits
27 Ready-to-deploy protection
29 Research overview
29 Introduction: Why obfuscation matters
29 What was tested: Top AI models vs. protected code
30 The test program
30 The four-dimensional framework: A new way to understand
AI capabilities
32 Results: How the AI models performed
33 Key findings: What this means for software protection
34 The three-tier resistance model: A practical guide
35 Common AI errors in code analysis
35 Implications and recommendations
36 Conclusion
Contents
Promon App Threat Report Q2 2025 2
4
Introduction
mobile application security conducted by our Security Research
Promon App Threat Report Q2 2025
threat landscape
5Promon App Threat Report Q2 2025
•
•
•
Europe
Germany
A national report called
The report showed that botnets were primarily used to steal
6
Promon App Threat Report Q2 2025
7Promon App Threat Report Q2 2025
The EU and beyond
This report is called the
and compliments the more general
Fraud: 6%
Primary cybersecurity threats
Social engineering
Ransomware
Fraud
Promon App Threat Report Q2 2025
•
•
•
•
•
• Emerging threats include:
•
•
•
and bypass traditional security measures
9Promon App Threat Report Q2 2025
•
The United States of America
The
•
•
•
•
A research report on
•
•
•
•
•
Promon App Threat Report Q2 2025
•
•
Southeast Asia
A report on
•
•
•
•
•
Promon App Threat Report Q2 2025
Promon App Threat Report Q2 2025
Fraud classification:
Fraud technique:
Malware type:
BingoMod
Promon App Threat Report Q2 2025
ToxicPanda
Android malware
Coper
Promon App Threat Report Q2 2025
Vultur
Promon App Threat Report Q2 2025
Promon App Threat Report Q2 2025
has warned
Promon App Threat Report Q2 2025
Deep fake threats in biometrics
Promon App Threat Report Q2 2025
Why AI-only solutions fall short in mobile deepfake
video detection
The computational reality of mobile devices
• Limited processing power:
• Battery impact:
• Memory constraints:
processes
• Heat generation:
As mentioned in our blog post on
The moving target problem
• Model staleness:
• Retraining requirements:
• Signature-based limitations:
Promon App Threat Report Q2 2025
These papers are:
Cross-platform vulnerabilities
• Platform optimization:
• Hardware diversity:
• API inconsistencies:
blind spots in detection
The case for layered defense
incorporates multiple complementary methods that operate at
Short-term defensive measures
• App shielding:
• KYC and identity verification:
20Promon App Threat Report Q2 2025
• Behavioral analysis:
• Transaction authentication: Implementing strong
Long-term strategic initiatives
• Platform-provided biometric APIs:
• Regular security testing:
• OS-level biometric security:
Conclusion: Beyond AI alone
mobile applications
Promon App Threat Report Q2 2025
22Promon App Threat Report Q2 2025
categories in action:
Runtime model tampering: Research demonstrates that
Local data store compromise:
AI agent runtime exploitation:
Prompt injection attacks:
Financial impact shows urgent need for protection
23Promon App Threat Report Q2 2025
Key takeaways
•
traditional application security concerns
•
•
• Implementing AI security measures is increasingly
The AI security challenge
• Unique assets:
• New attack surfaces:
• Complex interactions:
with multiple data sources and components in ways
that create security gaps
• Regulatory requirements: Emerging regulations
Critical AI threats on devices
•
•
•
24Promon App Threat Report Q2 2025
Runtime model tampering
What it is:
Business impact:
•
• Competitors gaining access to proprietary
AI capabilities
•
to incorrect business decisions
•
Real-world example: A competitor gains access to your
Local data store compromise
What it is:
Business impact:
•
•
•
•
Real-world example:
AI agent runtime exploitation
What it is:
Business impact:
•
•
•
unauthorized actions
25Promon App Threat Report Q2 2025
•
Real-world example:
Prompt injection attacks
What it is:
Business impact:
•
your content policies
•
in the model
•
•
Real-world example:
How can your AI applications be protected?
Runtime application shielding
Protects against:
with system prompts
Ready-to-deploy features:
•
•
•
remain unaltered during operation
Business benefit:
26Promon App Threat Report Q2 2025
Secure communications protection
Protects against:
Ready-to-deploy features:
• Hardened encrypted channels that secure
•
•
communication between app components and AI models
Business benefit:
A local data protection system
Protects against: Unauthorized access to internal data
sources used by AI
Ready-to-deploy features:
•
including models and system prompts
•
•
any tampering with stored AI data
Business benefit:
Regulatory compliance benefits
• EU AI Act:
• GDPR:
• Sector-specific regulations:
• Transparency requirements: Enables better
Promon’s protection solutions are available now to secure
your AI applications against these threats. Contact us today
for a demonstration of how our technology can protect your
specific AI deployments.
27Promon App Threat Report Q2 2025
Ready-to-deploy protection
As AI becomes central to business operations and customer
protection capabilities directly address the most critical threats
organization can:
Safeguard valuable AI intellectual property that
Strengthen the reliability and accuracy of AI-driven
decisions that your business depends on
Protect sensitive data processed by AI systems
Achieve compliance with emerging AI regulations
Prevent reputational damage from AI security
incidents that could undermine customer trust
Promon App Threat Report Q2 2025
29Promon App Threat Report Q2 2025
Research overview
Introduction: Why obfuscation matters
Legitimate use:
Malicious use:
Understanding how modern AI can analyze protected code helps
What was tested: Top AI models vs. protected code
This section presents original research conducted by
members of Promon’s Security Research Team.
The research was published in the paper ‘Deconstructing
30Promon App Threat Report Q2 2025
•
•
•
•
•
•
•
•
The test program
The researchers used a deliberately simple but strategically
Bogus control flow: Adding misleading decision paths
Instruction substitution: Replacing simple operations
Control flow flattening:
structure to hide its true operation
Combined techniques: Applying all three methods together
The four-dimensional framework: A new way to understand
AI capabilities
Reasoning depth:
Pattern recognition:
Promon App Threat Report Q2 2025
Noise filtering:
Context integration:
Results: How the AI models performed
Table 1 below
32Promon App Threat Report Q2 2025
Model Bogus control
flow
Instruction
substitution
Control flow
flattening
Combined
techniques
44
4 0 5
30 5
55 5
4 0 5
Sonnet 0 5
Table 1:
as the AI model produced no
scale:
• Level 0:
• Level 1-2:
• Level 3-4:
• Level 5:
correction
Level 5
Level 3-4
Level 1-2
Level 0
33Promon App Threat Report Q2 2025
Key findings: What this means for software protection
Significant variation between AI models
Some protection methods are already vulnerable
Instruction substitution remains highly effective
Layered protection works
AI capabilities are uneven and predictable
34Promon App Threat Report Q2 2025
The three-tier resistance model: A practical guide
Low resistance techniques
Moderate resistance techniques
High resistance techniques
AI pattern recognition or simultaneously challenge multiple
LOW
MODERATE
HIGH
Resistance AI capabilities
challenged
Level of human
assistance required
Example
Reasoning depth
Pattern recognition and
High
in current AI pattern
recognition or multiple
capability dimensions
simultaneously
or beyond correction
35Promon App Threat Report Q2 2025
Common AI errors in code analysis
made when analyzing protected code:
• Predicate misinterpretation: Failing to recognize that
• Structural mapping:
but incorrectly connecting them to the control structure
• Control flow misinterpretation: Incorrectly reconstructing
• Arithmetic transformation: Failing to correctly reconstruct
• Constant propagation:
Implications and recommendations
For software developers
• Layer your defenses:
• Focus on instruction substitution:
• Don’t rely solely on bogus control flow:
For security teams
• Use AI tools as supplements: Current AI models can reduce
• Combine AI strengths:
• Expect capabilities to evolve:
so protection and analysis strategies should be regularly
36Promon App Threat Report Q2 2025
Conclusion
opportunity to:
Build automated assessment systems that can systematically
as they emerge
Strategically improve protection mechanisms based on
strategies
dimensions simultaneously
to resist emerging AI capabilities
Prepare for a changing security landscape where the
balance between protection and analysis will continually
Financial app analysis
Security Research Team
37Promon App Threat Report Q2 2025
Promon App Threat Report Q2 2025
Methodology overview
Sample selection
Test process
•
•
•
Result omissions
Promon’s Security Research Team tested apps for this part
of the App Threat Report.
39Promon App Threat Report Q2 2025
When the app was not available in their region: This is a
When the app required a SIM card:
Testing summary
These were the results:
•
•
•
•
•
Key findings
•
•
•
• User warned: 6
•
•
40Promon App Threat Report Q2 2025
Key findings from tested finance apps
User warnedFully protected apps
The number of
vulnerable apps
is down 32.94%
from 2024
The number of
fully protected apps
is up 7.14%
from 2024
Conclusion
Promon App Threat Report Q2 2025
42Promon App Threat Report Q2 2025
mitigations/
Citations
Promon AS
promon.io
