🛡 AgentShield

Secure Agentic Browser Framework

Hack IITK 2026 – C3iHub Submission

---

Project Overview

AgentShield is a modular security framework designed to protect AI powered agentic browsers from malicious web interactions.

This project is developed as a working prototype for the Hack IITK 2026 Cybersecurity Hackathon organized by C3iHub, IIT Kanpur, under the problem statement:

“Securing Agentic Browsers Against Malicious Web Interactions.”

AgentShield introduces a layered defense architecture that detects adversarial web content, evaluates risk using weighted scoring, validates instructions against the agent’s declared objective, and enforces policy-based action mediation before browser execution.

---

Problem Context

AI driven browser agents are increasingly used for:

• Automated data extraction
• Form filling
• Ticket booking
• Online task automation

However, web environments are adversarial. Malicious pages may contain:

• Prompt injection attacks
• Hidden instructions via CSS
• Deceptive UI elements
• Phishing forms
• Dynamically injected scripts
• Instruction hijacking attempts

Traditional browser security mechanisms are designed for humans — not autonomous AI agents.

AgentShield addresses this gap.

---

System Architecture

Agent → Security Layer → Browser

The security layer intercepts and validates all web interactions before execution.

Security Modules

• DOM Parser`
• Prompt Injection Detector`
• Phishing Detector`
• UI Deception Analyzer`
• Dynamic Script Analyzer`
• Goal Consistency Validator`
• Weighted Risk Engine`
• Policy Based Decision Engine`
• Secure Action Executor`

---

Key Features

1️⃣ Multi Vector Threat Detection

Detects:

• Visible prompt injection
• Hidden CSS based injection
• Phishing style forms
• Suspicious JavaScript patterns
• Deceptive UI components

---

2️⃣ Goal Aware Validation Layer

Extracted instructions are compared against the agent’s declared objective.

This prevents:

• Instruction hijacking
• Task override attacks
• Unauthorized goal manipulation

---

3️⃣ Weighted Risk Scoring Engine

Each threat category contributes to a final risk score (0–100%).

Threat Type	Weight
Injection	30
Phishing	40
UI Deception	25
Dynamic Script	20
Goal Mismatch	35

Final risk score is capped at 100%.

---

4️⃣ Policy Based Secure Action Mediation

Before any browser action is executed:

• Risk ≥ 70 → BLOCK
• Risk ≥ 40 → CONFIRM
• Risk < 40 → ALLOW

All agent actions are intercepted and validated before execution.

---

5️⃣ Explainable Risk Assessment

For every interaction, AgentShield provides:

• Primary threat classification
• Detailed risk breakdown
• Evidence of detected threats
• Performance metrics

This ensures transparency and interpretability.

---

Demonstrated Attack Scenarios

The prototype defends against:

• Visible Prompt Injection
• Hidden CSS Based Injection
• Phishing Login Forms
• Instruction Goal Mismatch Attacks
• Safe Baseline Validation

---

Evaluation Metrics

The system demonstrates:

• High detection accuracy across multiple attack vectors
• Low false positive rate on safe content
• Real-time performance (~0.001s detection time)
• Clear and explainable mitigation decisions
• Secure action interception before browser execution

---

Performance

• Average Detection Time: ~0.001 seconds
• Minimal latency overhead
• Real time responsiveness

---

▶ How to Run

pip install -r requirements.txt
playwright install
python main.py

Environment Setup (For Judges)

To ensure smooth execution and avoid dependency issues, please use:

Recommended Python Version: Python 3.11

The project has been fully tested on Python 3.11.
Newer versions such as Python 3.12 or 3.13 may cause compatibility issues with Playwright dependencies.

---

🔎 Step 0: Verify Python Version

Check your installed Python version:

python --version

---

Alignment with Hack IITK Evaluation Criteria

AgentShield satisfies:

Malicious Web Content Detection

• Detects visible and hidden injection
• Identifies deceptive UI elements
• Analyzes DOM and script behavior

Secure Action Mediation

• Intercepts agent actions
• Enforces policy based execution
• Prevents unsafe browser actions

Explainable Risk Assessment

• Generates weighted risk scores
• Provides human readable threat evidence
• Displays transparent decision logic

Performance and Latency

• Real time detection
• Minimal overhead
• Efficient DOM parsing

---

Hack IITK Submission

This project is submitted as part of:

Hack IITK 2026 – Cybersecurity Hackathon organized by C3iHub, IIT Kanpur

AgentShield demonstrates a practical, modular, and explainable framework for securing agentic browser systems against adversarial web environments.

---

Future Improvements

• LLM assisted intent reasoning
• Dynamic DOM mutation monitoring
• Advanced clickjacking detection
• Scalable deployment architecture
• Browser extension integration