Jailbreak Detection

User Input
    ↓
[Stage 1: Pattern Matching] → Known attack signatures
    ↓
[Stage 2: ML Classifier]   → Learned jailbreak patterns
    ↓
[Stage 3: Perplexity Check] → Unusual token distributions
    ↓
[Stage 4: Context Analysis] → Multi-turn behavior patterns
    ↓
LLM Processing
    ↓
[Stage 5: Output Validation] → Response safety check
    ↓
User Response

import re
from typing import List, Tuple

class JailbreakPatternDetector:
    def __init__(self):
        self.patterns = [
            # DAN-style attacks
            (r"do anything now|DAN mode", "dan_attack", 0.9),
            (r"pretend you (are|have) no (restrictions|rules)",
             "persona_override", 0.85),
            # Role-play exploits
            (r"act as .* (without|no) (restrictions|limits)",
             "roleplay_exploit", 0.8),
            # Authority claims
            (r"(as|i am) (your|the|an?) (developer|creator|admin)",
             "authority_claim", 0.7),
            # Encoding requests
            (r"(decode|translate) (this )?(base64|rot13)",
             "encoding_bypass", 0.75),
        ]

    def detect(self, text: str) -> List[Tuple[str, float]]:
        results = []
        text_lower = text.lower()
        for pattern, label, confidence in self.patterns:
            if re.search(pattern, text_lower):
                results.append((label, confidence))
        return results

from transformers import pipeline

class JailbreakClassifier:
    def __init__(self, model_name="jailbreak-detector-v1"):
        self.classifier = pipeline(
            "text-classification",
            model=model_name
        )
        self.threshold = 0.7

    def is_jailbreak(self, text: str) -> dict:
        result = self.classifier(text)[0]
        return {
            "is_jailbreak": result["label"] == "JAILBREAK"
                            and result["score"] > self.threshold,
            "confidence": result["score"],
            "label": result["label"]
        }