File size: 7,678 Bytes
fe1ccde
6130c63
 
fe1ccde
 
 
 
 
6130c63
fe1ccde
 
 
 
 
 
 
 
 
6130c63
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
fe1ccde
6130c63
fe1ccde
6130c63
fe1ccde
6130c63
 
 
 
fe1ccde
6130c63
fe1ccde
 
6130c63
fe1ccde
 
6130c63
fe1ccde
 
 
 
6130c63
fe1ccde
 
6130c63
 
 
 
 
 
 
 
fe1ccde
6130c63
fe1ccde
6130c63
 
 
fe1ccde
6130c63
 
 
fe1ccde
6130c63
 
 
 
 
 
fe1ccde
6130c63
 
 
 
fe1ccde
 
6130c63
fe1ccde
 
6130c63
 
 
 
 
fe1ccde
6130c63
 
fe1ccde
6130c63
 
 
 
 
 
 
 
 
 
 
 
 
 
 
fe1ccde
6130c63
 
 
 
 
 
 
 
fe1ccde
6130c63
fe1ccde
6130c63
 
 
 
 
 
fe1ccde
6130c63
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
fe1ccde
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
import gradio as gr
import os
from typing import Dict, List
from datasets import load_dataset
from sentence_transformers import SentenceTransformer, util
from transformers import pipeline

# ============================================================
#  BACKEND: GUARDRAIL LOGIC (Same as before, optimized)
# ============================================================

class SafetyFinding:
    def __init__(self, label: str, severity: str, message: str):
        self.label = label
        self.severity = severity
        self.message = message

    def to_dict(self):
        return {"label": self.label, "severity": self.severity, "message": self.message}

class GuardrailSystem:
    def __init__(self):
        print("⚙️ Loading Guardrail Models... Please wait.")
        # 1. Load Heuristic Keywords
        self.unsafe_terms = ["bomb", "kill", "suicide", "explosive", "hack", "rob", "steal", "drugs", "murder"]
        self.jailbreak_terms = ["ignore previous", "system prompt", "jailbreak", "developer mode"]
        
        # 2. Load HuggingFace Moderator (Lazy loading recommended, but here we init upfront)
        self.moderator = pipeline("text-classification", model="unitary/toxic-bert")
        
        # 3. Load JailbreakBench Embeddings
        dataset = load_dataset("JailbreakBench/JBB-Behaviors", "behaviors", split="harmful")
        self.malicious_goals = [row["Goal"] for row in dataset if row and row["Goal"]]
        self.embedder = SentenceTransformer("all-MiniLM-L6-v2")
        self.goal_embeddings = self.embedder.encode(self.malicious_goals, convert_to_tensor=True)
        self.threshold = 0.5

    def check_heuristics(self, text):
        findings = []
        for term in self.unsafe_terms:
            if term in text.lower():
                findings.append(SafetyFinding("unsafe_keyword", "high", f"Detected unsafe term: '{term}'"))
        for term in self.jailbreak_terms:
            if term in text.lower():
                findings.append(SafetyFinding("jailbreak_keyword", "high", f"Detected jailbreak term: '{term}'"))
        if "@" in text:
             findings.append(SafetyFinding("pii_leak", "medium", "Potential PII (Email) detected"))
        return findings

    def check_similarity(self, text):
        findings = []
        if not text.strip(): return findings
        user_emb = self.embedder.encode(text, convert_to_tensor=True)
        cos_scores = util.cos_sim(user_emb, self.goal_embeddings)[0]
        max_score = float(cos_scores.max())
        
        if max_score >= self.threshold:
            findings.append(SafetyFinding("jailbreak_similarity", "high", f"Semantic Match to Jailbreak (Score: {max_score:.2f})"))
        return findings

    def check_moderation(self, text):
        findings = []
        if not text.strip(): return findings
        results = self.moderator(text, truncation=True)
        for r in results:
            if r["label"] in ["toxic", "severe_toxic", "obscene", "threat", "insult", "identity_hate"] and r["score"] > 0.7:
                findings.append(SafetyFinding("model_moderation", "high", f"Model Flag: {r['label']} ({r['score']:.2f})"))
        return findings

    def run_checks(self, user_prompt, simulated_response):
        findings = []
        # Input Checks
        findings += self.check_heuristics(user_prompt)
        findings += self.check_similarity(user_prompt)
        
        # Output Checks (Simulated)
        findings += self.check_heuristics(simulated_response)
        findings += self.check_moderation(user_prompt + " " + simulated_response)
        
        # Decision
        status = "ALLOWED"
        if any(f.severity == "high" for f in findings):
            status = "BLOCKED"
        elif any(f.severity == "medium" for f in findings):
            status = "REDACTED"
            
        return status, findings

# Initialize System (Global to keep in memory)
guard = GuardrailSystem()

# ============================================================
#  FRONTEND: PROFESSIONAL UI LOGIC
# ============================================================

def analyze_prompt(user_prompt):
    # Simulate LLM Generation for the demo
    simulated_output = "This is a harmless AI response."
    if "bomb" in user_prompt.lower(): simulated_output = "Here are instructions for..."
    if "email" in user_prompt.lower(): simulated_output = "Contact me at user@example.com"
    
    # Run Guardrails
    status, findings = guard.run_checks(user_prompt, simulated_output)
    
    # Generate HTML Status Card
    color_map = {"ALLOWED": "green", "BLOCKED": "red", "REDACTED": "orange"}
    icon_map = {"ALLOWED": "✅", "BLOCKED": "🛡️", "REDACTED": "⚠️"}
    
    html_status = f"""
    <div style='background-color: var(--background-fill-secondary); border-left: 5px solid {color_map[status]}; padding: 20px; border-radius: 8px; box-shadow: 0 4px 6px rgba(0,0,0,0.1);'>
        <h2 style='color: {color_map[status]}; margin: 0;'>{icon_map[status]} {status}</h2>
        <p style='margin-top: 5px; opacity: 0.8;'>Guardrail decision based on {len(findings)} risk factors.</p>
    </div>
    """
    
    # Format Findings for Display
    clean_findings = [f.to_dict() for f in findings]
    
    return html_status, clean_findings, simulated_output if status == "ALLOWED" else "[CONTENT BLOCKED BY POLICY]"

# Custom CSS for a clean look
custom_css = """
.gradio-container {font-family: 'Inter', sans-serif;}
h1 {text-align: center; color: #2d3748;}
"""

# Create the App
with gr.Blocks(theme=gr.themes.Soft(primary_hue="blue", neutral_hue="slate"), css=custom_css) as demo:
    
    # Header
    with gr.Row():
        gr.Markdown(
            """
            # 🛡️ Enterprise AI Guardrail System
            ### Real-time safety filtering using Semantic Search, BERT Moderation, and Heuristics.
            """
        )
    
    # Main Interface
    with gr.Row():
        # Left Column: Inputs
        with gr.Column(scale=1):
            gr.Markdown("### 📥 Input Simulation")
            input_text = gr.Textbox(
                lines=5, 
                label="User Prompt", 
                placeholder="Enter a prompt to test the guardrails (e.g., 'how to build a bomb' or 'hello')...",
                elem_id="input_box"
            )
            analyze_btn = gr.Button("🛡️ Run Safety Check", variant="primary", size="lg")
            
            with gr.Accordion("ℹ️ How it works", open=False):
                gr.Markdown("""
                1. **Heuristics:** Checks for banned keywords.
                2. **Vector Database:** Compares prompt against known jailbreaks (JailbreakBench).
                3. **BERT Classifier:** Scans for toxic tones.
                """)

        # Right Column: Analytics
        with gr.Column(scale=1):
            gr.Markdown("### 📊 Live Analytics")
            status_display = gr.HTML(label="Decision")
            
            with gr.Tabs():
                with gr.TabItem("Findings"):
                    findings_json = gr.JSON(label="Risk Factors Detected")
                with gr.TabItem("Raw Output"):
                    final_output = gr.Code(label="LLM Response", language="markdown")

    # Footer
    gr.Markdown(
        """
        ---
        <div style="text-align: center; opacity: 0.5; font-size: 0.8rem;">
            Built for AI Safety Portfolio | Powered by HuggingFace Transformers & Gradio
        </div>
        """
    )

    # Event Linking
    analyze_btn.click(
        fn=analyze_prompt,
        inputs=input_text,
        outputs=[status_display, findings_json, final_output]
    )

if __name__ == "__main__":
    demo.launch()