AI Model Xavfsizligi: Adversarial Hujumlar va Red Teaming

Neyron Tarmoqlarga Hujum Turlari: MITRE ATLAS Matritsasi

FGSM Formulasi: Qanday Qilib 1% Shovqin Modelni Aldaydi?

Data Poisoning: O'quv Ma'lumotlarini Yashirincha Zaharlash

Prompt Injection: To'g'ridan-To'g'ri va Bilvosita Hujumlar

NeMo Guardrails va Llama Guard: AI Xavfsizlik Devorlari

Python da Ko'p Qatlamli Guardrail Tizimini Qurish

import re
import math

class EnterpriseAIGuardrail:
    def __init__(self, canary_token="CANARY_7X9Q"):
        self.canary = canary_token
        # Regex heuristics for known adversarial jailbreak patterns
        self.jailbreak_patterns = [
            re.compile(r"(?i)(ignore previous instructions|disregard all rules)"),
            re.compile(r"(?i)(you are now dan|unfiltered mode enabled)"),
            re.compile(r"(?i)(repeat the system prompt|reveal secret instructions)"),
        ]

    def validate_input(self, user_prompt: str) -> bool:
        # 1. Heuristic Pattern Matching
        for pattern in self.jailbreak_patterns:
            if pattern.search(user_prompt):
                raise ValueError("Security Alert: Prompt Injection Pattern Detected!")
        
        # 2. Shannon Entropy Analysis (Detects Base64 / Obfuscated ciphers)
        prob = [float(user_prompt.count(c)) / len(user_prompt) for c in set(user_prompt)]
        entropy = -sum(p * math.log2(p) for p in prob) if prob else 0
        if len(user_prompt) > 50 and entropy > 4.8:
            raise ValueError("Security Alert: Abnormal high-entropy payload detected (Obfuscation)!")
        return True

    def validate_output(self, model_response: str) -> str:
        # Check for system canary token exfiltration
        if self.canary in model_response:
            raise SecurityError("Critical Breach: System Prompt Leakage Prevented!")
        # Redact credit card numbers
        redacted = re.sub(r'\b(?:\d[ -]*?){13,16}\b', '[REDACTED_CARD]', model_response)
        return redacted

Chevrolet Dilerlik Falokati va Sydney Maxfiy Prompti Sizishi

AI Red Teaming: Modellarni Chiqarishdan Oldin Buzib Ko'rish

Amaliy Ish: AI Red Teaming Studio — FGSM va Guardrail Kvestlari

AI Xavfsizlik Qatlamlarining Taqqoslama Matritsasi

Xulosa va Uy Vazifasi: Korporativ LLM Xavfsizlik Nizomi