Chapitre 3 : Tests de sécurité

Red Teaming

Le red teaming consiste à essayer délibérément de faire produire à l'LLM des sorties nuisibles.

class RedTeamer:
    def __init__(self, llm):
        self.llm = llm
        self.attack_categories = [
            "jailbreak_attempts",
            "prompt_injection",
            "harmful_content_elicitation",
            "pii_extraction",
            "bias_probing"
        ]
    
    def run_attack_suite(self, category: str) -> list:
        attacks = self.load_attacks(category)
        results = []
        
        for attack in attacks:
            response = self.llm.generate(attack["prompt"])
            is_safe = self.evaluate_safety(response)
            
            results.append({
                "attack": attack["name"],
                "category": category,
                "response": response[:200],
                "is_safe": is_safe
            })
        
        return results

Référentiels de sécurité

Référentiel	Focus	Mesures
TruthfulQA	Véracité	Précision MC
ToxiGen	Toxicité	Taux de toxicité
BBQ	Biais social	Score de biais
HarmBench	Comportements nuisibles	Taux de réussite des attaques

Pipeline de filtrage de contenu

class ContentFilter:
    def __init__(self):
        self.toxicity_model = load_model("unitary/toxic-bert")
        
    def filter(self, text: str) -> dict:
        # Check toxicity
        toxicity_score = self.toxicity_model.predict(text)
        
        # Check for PII
        pii_found = self.detect_pii(text)
        
        # Check for blocked patterns
        blocked = self.check_blocked_patterns(text)
        
        return {
            "text": text,
            "is_safe": toxicity_score < 0.5 and not pii_found and not blocked,
            "toxicity": toxicity_score,
            "pii_detected": pii_found,
            "blocked_content": blocked
        }

---

Suivant : Quiz Module 12