Securite et Conformite RAG : RGPD, AI Act et bonnes pratiques

Deployer un systeme RAG en entreprise implique de manipuler des donnees souvent sensibles : documents internes, informations clients, propriete intellectuelle. Ce guide couvre les aspects securite et conformite essentiels pour un deploiement serein.

Le cadre reglementaire europeen

RGPD : les fondamentaux pour le RAG

Le Reglement General sur la Protection des Donnees s'applique des qu'un systeme RAG traite des donnees personnelles europeennes.

AI Act : ce qui change pour les systemes RAG

L'AI Act europeen (entree en vigueur progressive 2024-2027) classifie les systemes IA par niveau de risque :

Risque limite (majorite des RAG) :

• Chatbots support client
• Assistants internes
• Recherche documentaire

Obligations :

• Transparence : indiquer que l'utilisateur interagit avec une IA
• Conservation des logs pour audit

Risque eleve (cas specifiques) :

• RAG pour decisions RH (recrutement, evaluation)
• RAG pour decisions credit/assurance
• RAG sante avec impact medical

Obligations supplementaires :

• Evaluation de conformite
• Systeme de gestion des risques
• Donnees d'entrainement documentees
• Supervision humaine

DEVELOPERpython
# Classification du risque selon l'AI Act
def classify_rag_risk(use_case: dict) -> str:
    high_risk_domains = [
        "hr_recruitment",
        "hr_evaluation",
        "credit_scoring",
        "insurance_pricing",
        "medical_diagnosis",
        "legal_decision"
    ]

    if use_case.get("domain") in high_risk_domains:
        return "high_risk"

    if use_case.get("automated_decision") and use_case.get("significant_impact"):
        return "high_risk"

    return "limited_risk"

Architecture securisee

Separation des environnements

┌─────────────────────────────────────────────────────────────┐
│                     PRODUCTION                               │
│  ┌─────────────┐  ┌─────────────┐  ┌─────────────────────┐  │
│  │   API GW    │──│   RAG API   │──│   Vector DB         │  │
│  │   (WAF)     │  │  (Isolated) │  │   (Encrypted)       │  │
│  └─────────────┘  └─────────────┘  └─────────────────────┘  │
│         │                │                   │               │
│         │                │                   │               │
│  ┌─────────────────────────────────────────────────────┐    │
│  │              Network Isolation (VPC)                 │    │
│  └─────────────────────────────────────────────────────┘    │
└─────────────────────────────────────────────────────────────┘

┌─────────────────────────────────────────────────────────────┐
│                      STAGING                                 │
│           (Donnees anonymisees uniquement)                  │
└─────────────────────────────────────────────────────────────┘

┌─────────────────────────────────────────────────────────────┐
│                    DEVELOPMENT                               │
│             (Donnees synthetiques)                          │
└─────────────────────────────────────────────────────────────┘

Chiffrement bout en bout

DEVELOPERpython
from cryptography.fernet import Fernet
from cryptography.hazmat.primitives import hashes
from cryptography.hazmat.primitives.kdf.pbkdf2 import PBKDF2HMAC
import base64

class SecureRAGStorage:
    def __init__(self, master_key: bytes):
        self.cipher = Fernet(master_key)

    def encrypt_document(self, content: str, metadata: dict) -> dict:
        """Chiffre le contenu avant stockage"""
        encrypted_content = self.cipher.encrypt(content.encode())

        # Chiffrer les metadonnees sensibles
        sensitive_fields = ["author", "email", "department"]
        encrypted_metadata = metadata.copy()

        for field in sensitive_fields:
            if field in encrypted_metadata:
                encrypted_metadata[field] = self.cipher.encrypt(
                    str(encrypted_metadata[field]).encode()
                ).decode()

        return {
            "content": encrypted_content.decode(),
            "metadata": encrypted_metadata,
            "encrypted": True
        }

    def decrypt_document(self, encrypted_doc: dict) -> dict:
        """Dechiffre le document pour utilisation"""
        if not encrypted_doc.get("encrypted"):
            return encrypted_doc

        content = self.cipher.decrypt(
            encrypted_doc["content"].encode()
        ).decode()

        return {
            "content": content,
            "metadata": self._decrypt_metadata(encrypted_doc["metadata"])
        }

    def encrypt_embedding(self, embedding: list[float]) -> bytes:
        """Chiffre les embeddings (optionnel, impact performance)"""
        import numpy as np
        embedding_bytes = np.array(embedding).tobytes()
        return self.cipher.encrypt(embedding_bytes)

Controle d'acces granulaire (RBAC)

DEVELOPERpython
from enum import Enum
from functools import wraps

class Permission(Enum):
    READ_PUBLIC = "read:public"
    READ_INTERNAL = "read:internal"
    READ_CONFIDENTIAL = "read:confidential"
    READ_RESTRICTED = "read:restricted"
    ADMIN = "admin"

class DocumentClassification(Enum):
    PUBLIC = "public"
    INTERNAL = "internal"
    CONFIDENTIAL = "confidential"
    RESTRICTED = "restricted"

class RBACController:
    def __init__(self, user_service):
        self.user_service = user_service

        self.permission_hierarchy = {
            Permission.READ_PUBLIC: [DocumentClassification.PUBLIC],
            Permission.READ_INTERNAL: [
                DocumentClassification.PUBLIC,
                DocumentClassification.INTERNAL
            ],
            Permission.READ_CONFIDENTIAL: [
                DocumentClassification.PUBLIC,
                DocumentClassification.INTERNAL,
                DocumentClassification.CONFIDENTIAL
            ],
            Permission.READ_RESTRICTED: [
                DocumentClassification.PUBLIC,
                DocumentClassification.INTERNAL,
                DocumentClassification.CONFIDENTIAL,
                DocumentClassification.RESTRICTED
            ]
        }

    async def filter_accessible_documents(
        self,
        documents: list[dict],
        user_id: str
    ) -> list[dict]:
        """Filtre les documents selon les permissions utilisateur"""
        user = await self.user_service.get_user(user_id)
        user_permissions = set(user.get("permissions", []))

        accessible = []
        for doc in documents:
            doc_class = DocumentClassification(
                doc.get("metadata", {}).get("classification", "internal")
            )

            if self._can_access(user_permissions, doc_class):
                accessible.append(doc)

        return accessible

    def _can_access(
        self,
        user_permissions: set,
        doc_classification: DocumentClassification
    ) -> bool:
        """Verifie si l'utilisateur peut acceder au document"""
        for perm in user_permissions:
            try:
                allowed = self.permission_hierarchy.get(Permission(perm), [])
                if doc_classification in allowed:
                    return True
            except ValueError:
                continue
        return False

Protection des donnees personnelles

Detection et anonymisation automatique

DEVELOPERpython
import re
from presidio_analyzer import AnalyzerEngine
from presidio_anonymizer import AnonymizerEngine

class PIIProtector:
    def __init__(self):
        self.analyzer = AnalyzerEngine()
        self.anonymizer = AnonymizerEngine()

        # Patterns personnalises pour la France
        self.french_patterns = {
            "FRENCH_SSN": r"\b[12][0-9]{2}[0-1][0-9][0-9]{2}[0-9]{3}[0-9]{3}[0-9]{2}\b",
            "FRENCH_PHONE": r"\b(?:(?:\+33|0033|0)[1-9](?:[.\-\s]?\d{2}){4})\b",
            "IBAN_FR": r"\bFR\d{2}[\s]?\d{4}[\s]?\d{4}[\s]?\d{4}[\s]?\d{4}[\s]?\d{3}\b"
        }

    def detect_pii(self, text: str, language: str = "fr") -> list[dict]:
        """Detecte les donnees personnelles dans le texte"""
        # Detection Presidio (email, nom, adresse, etc.)
        results = self.analyzer.analyze(
            text=text,
            language=language,
            entities=[
                "PERSON", "EMAIL_ADDRESS", "PHONE_NUMBER",
                "LOCATION", "CREDIT_CARD", "IBAN_CODE"
            ]
        )

        # Detection patterns francais
        french_results = self._detect_french_patterns(text)

        all_results = list(results) + french_results
        return self._format_results(all_results)

    def anonymize(
        self,
        text: str,
        mode: str = "replace"  # replace, hash, mask
    ) -> str:
        """Anonymise les donnees personnelles"""
        results = self.analyzer.analyze(text=text, language="fr")

        if mode == "replace":
            anonymized = self.anonymizer.anonymize(
                text=text,
                analyzer_results=results
            )
        elif mode == "hash":
            anonymized = self._hash_pii(text, results)
        elif mode == "mask":
            anonymized = self._mask_pii(text, results)

        return anonymized.text

    def _hash_pii(self, text: str, results: list) -> str:
        """Remplace les PII par leur hash (reversible avec cle)"""
        import hashlib

        result_text = text
        for result in sorted(results, key=lambda x: x.start, reverse=True):
            original = text[result.start:result.end]
            hashed = hashlib.sha256(original.encode()).hexdigest()[:12]
            result_text = result_text[:result.start] + f"[{result.entity_type}:{hashed}]" + result_text[result.end:]

        return result_text

    def check_before_indexing(self, document: dict) -> dict:
        """Verifie et traite un document avant indexation"""
        content = document.get("content", "")

        pii_detected = self.detect_pii(content)

        if pii_detected:
            return {
                "safe_to_index": False,
                "pii_found": pii_detected,
                "action_required": "review_or_anonymize",
                "anonymized_content": self.anonymize(content)
            }

        return {
            "safe_to_index": True,
            "pii_found": [],
            "original_content": content
        }

Consentement et droits des personnes

DEVELOPERpython
class ConsentManager:
    def __init__(self, db):
        self.db = db

    async def record_consent(
        self,
        user_id: str,
        purpose: str,
        data_categories: list[str],
        expiry_date: datetime = None
    ):
        """Enregistre le consentement utilisateur"""
        consent = {
            "user_id": user_id,
            "purpose": purpose,
            "data_categories": data_categories,
            "granted_at": datetime.now(),
            "expires_at": expiry_date,
            "status": "active"
        }

        await self.db.insert("consents", consent)

    async def check_consent(
        self,
        user_id: str,
        purpose: str,
        data_category: str
    ) -> bool:
        """Verifie si le consentement est valide"""
        consent = await self.db.find_one("consents", {
            "user_id": user_id,
            "purpose": purpose,
            "data_categories": {"$in": [data_category]},
            "status": "active",
            "$or": [
                {"expires_at": None},
                {"expires_at": {"$gt": datetime.now()}}
            ]
        })

        return consent is not None

    async def handle_deletion_request(self, user_id: str) -> dict:
        """Traite une demande de suppression (droit a l'oubli)"""
        # 1. Identifier toutes les donnees
        user_data = await self._find_all_user_data(user_id)

        # 2. Supprimer des index vectoriels
        await self._delete_from_vector_db(user_data["document_ids"])

        # 3. Supprimer des bases de donnees
        await self._delete_from_databases(user_id)

        # 4. Journaliser la suppression
        await self._log_deletion(user_id, user_data)

        return {
            "status": "completed",
            "deleted_documents": len(user_data["document_ids"]),
            "deleted_conversations": user_data["conversation_count"],
            "completion_date": datetime.now().isoformat()
        }

    async def handle_export_request(self, user_id: str) -> dict:
        """Traite une demande d'export (droit a la portabilite)"""
        user_data = await self._find_all_user_data(user_id)

        export = {
            "user_profile": user_data["profile"],
            "conversations": user_data["conversations"],
            "indexed_documents": user_data["documents"],
            "consents": user_data["consents"],
            "export_date": datetime.now().isoformat()
        }

        return export

Audit et tracabilite

Logging securise

DEVELOPERpython
import logging
import json
from datetime import datetime

class AuditLogger:
    def __init__(self, log_destination: str):
        self.logger = logging.getLogger("rag_audit")
        self.logger.setLevel(logging.INFO)

        # Handler securise (fichier chiffre ou SIEM)
        handler = logging.FileHandler(log_destination)
        handler.setFormatter(logging.Formatter(
            '%(asctime)s - %(levelname)s - %(message)s'
        ))
        self.logger.addHandler(handler)

    def log_query(
        self,
        user_id: str,
        query: str,
        documents_accessed: list[str],
        response_generated: bool
    ):
        """Log chaque requete RAG"""
        # NE PAS logger le contenu de la query si sensible
        log_entry = {
            "event_type": "rag_query",
            "timestamp": datetime.now().isoformat(),
            "user_id": self._hash_if_needed(user_id),
            "query_hash": self._hash_content(query),
            "query_length": len(query),
            "documents_accessed": documents_accessed,
            "document_count": len(documents_accessed),
            "response_generated": response_generated
        }

        self.logger.info(json.dumps(log_entry))

    def log_document_access(
        self,
        user_id: str,
        document_id: str,
        access_type: str,
        classification: str
    ):
        """Log les acces aux documents"""
        log_entry = {
            "event_type": "document_access",
            "timestamp": datetime.now().isoformat(),
            "user_id": self._hash_if_needed(user_id),
            "document_id": document_id,
            "access_type": access_type,
            "classification": classification
        }

        self.logger.info(json.dumps(log_entry))

    def log_security_event(
        self,
        event_type: str,
        severity: str,
        details: dict
    ):
        """Log les evenements securite"""
        log_entry = {
            "event_type": f"security_{event_type}",
            "timestamp": datetime.now().isoformat(),
            "severity": severity,
            "details": details
        }

        if severity in ["high", "critical"]:
            self.logger.warning(json.dumps(log_entry))
            self._alert_security_team(log_entry)
        else:
            self.logger.info(json.dumps(log_entry))

    def _hash_content(self, content: str) -> str:
        import hashlib
        return hashlib.sha256(content.encode()).hexdigest()

Retention et purge

DEVELOPERpython
class DataRetentionManager:
    def __init__(self, db, vector_db, config: dict):
        self.db = db
        self.vector_db = vector_db
        self.config = config

    async def apply_retention_policy(self):
        """Applique la politique de retention"""

        # Conversations
        conv_retention = self.config.get("conversation_retention_days", 365)
        await self._purge_old_conversations(conv_retention)

        # Logs d'audit (obligation legale plus longue)
        audit_retention = self.config.get("audit_retention_days", 2190)  # 6 ans
        await self._archive_old_audits(audit_retention)

        # Documents expires
        await self._handle_expired_documents()

        # Metadonnees orphelines
        await self._cleanup_orphaned_data()

    async def _purge_old_conversations(self, days: int):
        """Supprime les conversations anciennes"""
        cutoff_date = datetime.now() - timedelta(days=days)

        # Anonymiser plutot que supprimer si besoin analytics
        await self.db.update_many(
            "conversations",
            {"created_at": {"$lt": cutoff_date}},
            {
                "$set": {
                    "user_id": "anonymized",
                    "messages": [],
                    "purged_at": datetime.now()
                }
            }
        )

    async def _handle_expired_documents(self):
        """Gere les documents avec date d'expiration"""
        expired = await self.db.find(
            "documents",
            {"expires_at": {"$lt": datetime.now()}}
        )

        for doc in expired:
            # Supprimer de l'index vectoriel
            await self.vector_db.delete(doc["id"])

            # Marquer comme expire
            await self.db.update(
                "documents",
                {"id": doc["id"]},
                {"$set": {"status": "expired", "content": None}}
            )

Securite operationnelle

Protection contre les injections

DEVELOPERpython
class PromptSecurityGuard:
    def __init__(self):
        self.injection_patterns = [
            r"ignore\s+(previous|all|above)\s+instructions",
            r"disregard\s+(your|the)\s+(rules|instructions)",
            r"you\s+are\s+now\s+",
            r"pretend\s+(you|to)\s+",
            r"act\s+as\s+if",
            r"system\s*:\s*",
            r"<\|.*\|>",
            r"\[INST\]",
            r"###\s*(instruction|system)",
        ]

    def check_query(self, query: str) -> dict:
        """Verifie la securite d'une requete"""
        query_lower = query.lower()

        # Detection patterns d'injection
        for pattern in self.injection_patterns:
            if re.search(pattern, query_lower, re.IGNORECASE):
                return {
                    "safe": False,
                    "reason": "potential_injection",
                    "pattern_matched": pattern
                }

        # Detection tentatives d'exfiltration
        if self._detect_exfiltration_attempt(query):
            return {
                "safe": False,
                "reason": "potential_exfiltration"
            }

        # Detection contenu malveillant
        if self._detect_malicious_content(query):
            return {
                "safe": False,
                "reason": "malicious_content"
            }

        return {"safe": True}

    def sanitize_context(self, context: str) -> str:
        """Nettoie le contexte avant injection dans le prompt"""
        # Supprimer les tentatives d'injection dans les documents
        sanitized = context

        for pattern in self.injection_patterns:
            sanitized = re.sub(pattern, "[FILTERED]", sanitized, flags=re.IGNORECASE)

        return sanitized

    def _detect_exfiltration_attempt(self, query: str) -> bool:
        """Detecte les tentatives d'exfiltration de donnees"""
        exfil_patterns = [
            r"list\s+all\s+(users|passwords|keys|secrets)",
            r"show\s+(me\s+)?(the\s+)?system\s+prompt",
            r"what\s+(are|is)\s+(your|the)\s+(instructions|rules)",
            r"dump\s+(all|the)\s+data",
            r"export\s+.*\s+to\s+(email|url|external)"
        ]

        for pattern in exfil_patterns:
            if re.search(pattern, query.lower()):
                return True

        return False

Rate limiting et protection DDoS

DEVELOPERpython
from redis import Redis
import time

class RateLimiter:
    def __init__(self, redis_client: Redis):
        self.redis = redis_client

    async def check_rate_limit(
        self,
        identifier: str,
        limit_type: str = "query"
    ) -> dict:
        """Verifie les limites de taux"""

        limits = {
            "query": {"requests": 60, "window": 60},      # 60 req/min
            "heavy_query": {"requests": 10, "window": 60}, # 10 req/min
            "indexing": {"requests": 100, "window": 3600}, # 100/heure
            "export": {"requests": 5, "window": 86400}     # 5/jour
        }

        config = limits.get(limit_type, limits["query"])

        key = f"ratelimit:{limit_type}:{identifier}"
        current = await self.redis.incr(key)

        if current == 1:
            await self.redis.expire(key, config["window"])

        remaining = max(0, config["requests"] - current)

        if current > config["requests"]:
            return {
                "allowed": False,
                "remaining": 0,
                "reset_in": await self.redis.ttl(key)
            }

        return {
            "allowed": True,
            "remaining": remaining,
            "limit": config["requests"]
        }

Checklist conformite

Avant mise en production

DEVELOPERmarkdown
## RGPD
- [ ] Registre des traitements mis a jour
- [ ] Analyse d'impact (DPIA) si donnees sensibles
- [ ] Information des personnes concernees
- [ ] Mecanismes d'exercice des droits implementes
- [ ] Contrats sous-traitants (DPA) signes
- [ ] Mesures de securite documentees

## AI Act (si applicable)
- [ ] Classification du risque effectuee
- [ ] Systeme de gestion des risques (si haut risque)
- [ ] Documentation technique complete
- [ ] Mecanisme de supervision humaine
- [ ] Tests de robustesse et biais

## Securite
- [ ] Chiffrement au repos et en transit
- [ ] Controle d'acces RBAC configure
- [ ] Audit logging actif
- [ ] Tests de penetration effectues
- [ ] Plan de reponse aux incidents
- [ ] Sauvegardes testees

## Operationnel
- [ ] Politique de retention definie
- [ ] Processus de purge automatise
- [ ] Monitoring securite actif
- [ ] Formation equipes effectuee

Pour aller plus loin

• RGPD et Chatbots - Focus chatbot client
• AI Act et RAG - Implications reglementaires
• Donnees sensibles - Traitement donnees critiques

Conformite simplifiee avec Ailog

La conformite RGPD et AI Act peut sembler complexe. Avec Ailog, beneficiez d'une infrastructure deja conforme :

• Hebergement 100% France chez OVH, donnees jamais transferees hors UE
• Chiffrement bout en bout AES-256
• RBAC natif avec gestion des permissions par document
• Anonymisation automatique des donnees personnelles
• Logs d'audit exportables pour vos DPO
• DPA signe et documentation RGPD fournie

Testez Ailog gratuitement et deployez un RAG conforme des aujourd'hui.