Hello, World.

I'm SwiftGuard.

Fast Secure Intelligent

Learn More
About
Profile Picture

LLMs are powerful—but also vulnerable. SwiftGuard is the next-generation defense system that protects AI models from jailbreaking attacks without slowing them down.

Practice

  • Purpose: Protects AI models from jailbreaking attacks.
  • Core Technology: Single-Pass Detection (SPD) & Logit-Based Classification.
  • Use Cases: AI Security for Healthcare, Finance, and Enterprise AI.
  • Efficiency: Minimal computational overhead for seamless user experience.

COMPETENCIES

  • 88%
    Attack Detection Accuracy
  • 9.67%
    False Positive Rate
  • 85.48%
    F1 Score
  • 0.3s
    Average Response Time
Choose Your Story

The Right Protection for Your Needs

SwiftGuard offers two powerful configurations to match your specific requirements. Choose between speed and precision based on your use case.

SwiftGuard Classic

Our speed-optimized solution with higher detection rate

  • 88.08%
    Attack Detection
  • 9.67%
    False Positive Rate
  • 85.48%
    F1 Score
  • 0.3s
    Response Time

Ideal For:

  • Consumer applications
  • High-traffic environments
  • Real-time chat systems
  • Scenarios where detection rate and speed are critical

SwiftGuard Precision

Our reliability-optimized solution with minimal false positives

  • 86.5%
    Attack Detection
  • 0.5%
    False Positive Rate
  • 92.51%
    F1 Score
  • 4.62s
    Response Time

Ideal For:

  • Enterprise environments
  • Healthcare & financial systems
  • Scientific research applications
  • Scenarios where minimizing false positives is essential

Understanding the Tradeoff

The fundamental tradeoff in LLM protection systems is between speed and precision. SwiftGuard Classic processes prompts with minimal overhead, making it ideal for applications where user experience depends on rapid response times. SwiftGuard Precision incorporates an additional preliminary classifier that significantly reduces false positives at the cost of increased processing time, making it the preferred choice for environments where accuracy is the top priority.

Both configurations use the same core technology but optimize for different priorities, allowing you to select the right tool for your specific needs.

How It Works

Our Model Pipeline

SwiftGuard uses a two-stage filtering system to efficiently identify and block harmful prompts while allowing legitimate queries to pass through with minimal latency.

Benign Prompt Flow

User Prompt
Rule-Based Classifier
Preliminary Check
✓ Passes
Prompt Processed by LLM

Safe prompts flow through our system with minimal overhead, ensuring a seamless user experience.

Adversarial Prompt Flow

Harmful Prompt
Rule-Based Classifier
Suspicious Pattern
⚠ Flagged
Single-Pass Detection (SPD)
High Risk Score
❌ Harmful
Prompt Blocked & Logged

Malicious jailbreak attempts are identified with 88% accuracy and blocked before they can reach the underlying LLM, protecting the system from exploitation.

Why It Matters

Securing the Future of AI

As large language models become increasingly embedded in critical systems, protecting them from exploitation is not just a technical challenge—it's an ethical imperative.

The Stakes Are High

Without robust protection, LLMs can be manipulated to:

  • Generate harmful content despite safety guardrails
  • Leak sensitive information from training data
  • Execute potentially dangerous code or commands
  • Undermine trust in AI systems across sectors

SwiftGuard addresses these vulnerabilities head-on, providing a critical layer of defense without compromising the user experience.

Real-World Impact

Our exceptional results demonstrate SwiftGuard's potential to transform LLM security:

  • 88% detection accuracy identifies the vast majority of jailbreak attempts
  • Low 9.67% false positive rate ensures legitimate queries aren't blocked
  • Strong 85.48% F1 score validates our balanced approach to security
  • Ultra-fast 0.3s average response time maintains seamless user interaction

These metrics translate to safer, more reliable AI systems that organizations can deploy with confidence.

SwiftGuard represents a significant advancement in AI safety technology—balancing robust protection with computational efficiency to secure the next generation of language models.

About Us

Meet Our Team

SwiftGuard was developed as our capstone project at UC San Diego. Our team is passionate about AI security and dedicated to creating robust solutions that protect LLMs from adversarial attacks.

Contact

Contact Us!

Your message was sent, thank you!
Our Team

Shreya Sudan - Data Scientist
shreyasudan2211@gmail.com
LinkedIn | GitHub
Personal Website

Arman Rahman - ML Engineer
arahman@ucsd.edu
LinkedIn | GitHub

Donald Taggart - Data Analyst
dtaggart@ucsd.edu
LinkedIn | GitHub

Dante Testini - System Architect
dtestini@ucsd.edu
LinkedIn | GitHub

Faculty Mentors

Prof. Barna Saha
bsaha@ucsd.edu

Prof. Arya Mazumdar
arya@ucsd.edu