DEV Community

Jana Stalin
Jana Stalin

Posted on

CyberGuard: The AI Agent I Built to Fight Digital Scams (My AI Agents Intensive Journey)

Hey everyone,

I recently completed the Google × Kaggle AI Agents Intensive (5-Day course) — and this experience genuinely changed how I look at AI systems.

I joined because I kept seeing the term “AI Agents” everywhere, but honestly… I didn’t fully get it.

Before:
“Agents are basically chatbots with extra steps.”

Now:
Agents are systems that can think, plan, take actions, and collaborate with other agents to solve a task.

Here’s what I learned and how I built CyberGuard — a cybersecurity-focused AI agent system.

Day 1 — My Mental Shift

This was the biggest realization:

  • The LLM is the brain
  • Tools and services are the hands
  • The agent decides how to solve a task using those tools

It’s like working with a teammate who can:

  • Search
  • Run models
  • Make decisions
  • Remember what happened earlier

Once that clicked, I knew what I wanted to build.

The Problem I Wanted to Solve

Today, anyone can be tricked by:

  • Fake password reset emails
  • Apps that request dangerous permissions
  • Links pretending to be legitimate websites

Real attackers combine these:
Email scam leads to installing a malicious app → total account access.

Traditional security tools usually detect only one of these at a time.

That gap scared me enough to build something meaningful.

Introducing CyberGuard

A multi-agent system that examines:

Component Role Detection Scope Logic / Mechanism
Email Agent Phishing Analyzer Content & Links: Scans email bodies, sender metadata, and embedded URLs. Analyzes semantic patterns for social engineering cues and cross-references URLs with reputation blacklists to identify phishing vectors.
Android Agent Malware Scanner Permissions & Runtime: Inspects application behavior and access rights. Evaluates AndroidManifest.xml for over-privileged requests (e.g., SMS/Contact access) that indicate potential malware or spyware behavior.
Root Agent Risk Orchestrator Decision Engine: Aggregates and correlates signals. Synthesizes alerts from worker agents to determine a Combined Risk severity level, reducing false positives through cross-validation.

It returns a clear verdict:

  • SAFE
  • HIGH RISK
  • CRITICAL (extreme danger)

With an explanation someone can actually understand.

How I Built It (Simplified)

Phishing Detection Model

  • TF-IDF analysis + URL heuristic feature
  • Logistic Regression
  • Test performance: F1 Score ≈ 0.96
  • Served via FastAPI on port 8001

Malware Behavior Model

  • Random Forest Classifier on Android permission sets
  • Test performance: F1 Score ≈ 0.95
  • Served via FastAPI on port 8002

Both connect to a Root Agent running Gemini through the Agent Development Kit (ADK).

The Root Agent:

  • Calls models only when necessary
  • Correlates results from both services
  • Generates a Markdown “Security Audit” report with recommended actions

Example (shortened):

Severity: HIGH RISK  
Reason: Suspicious link in email + risky app permissions  
Recommendation: Delete the email and uninstall the app  
Enter fullscreen mode Exit fullscreen mode

Testing the Full System

I tested four scenarios:

Scenario Expected Result
Normal email + safe app Safe Safe
Phishing email only High High
Malware-like app only High High
Combination attack Critical Critical

Everything worked end-to-end.

It finally felt like a product, not just a notebook.

Challenges I Faced

  • Making different microservices communicate reliably
  • Debugging agents when their reasoning drifted
  • Matching permissions properly across datasets
  • Styling the UI took longer than expected

But every time a correct “CRITICAL” verdict appeared, it reminded me why I built this.

What’s Next

I’d like to:

  • Improve permission datasets for real Android apps
  • Integrate live threat-intelligence sources
  • Add browser extension to catch phishing instantly
  • Strengthen prompt security (agent safety is a real concern)

The system works, but there’s so much room to grow.

Final Thoughts

I started this course thinking that AI agents were too advanced or complex for me.

By Day 5, I had created a working cybersecurity assistant.

This wasn’t just about learning — it made me feel capable of building real, impactful AI systems.

If you’re curious about AI beyond chatbots:

Try building an agent.

It changes how you think.

Thank you for reading.

Happy to share the project or answer questions anytime.

Top comments (0)