AI Safety + Alignment: Jailbreak Defense, Red-Teaming, Constitutional AI, KVKK Uyumluluğu

Ders Haritası (10 Bölüm)#

1. AI safety neden — production'ın zorunluluğu
2. Jailbreak techniques — DAN, prompt injection
3. Defense layers — multi-stage protection
4. Red-teaming — internal adversarial testing
5. Constitutional AI (Bai 2022) — Anthropic'in yaklaşımı
6. OpenAI alignment stack — guidelines + RLHF + filtering
7. Content moderation — toxic classifier, output filtering
8. KVKK compliance — Türkiye veri koruma
9. AB AI Act 2024 — Europe regulation
10. Production safety checklist

2-7. Safety Techniques#

2.1 Jailbreak techniques#

• DAN (Do Anything Now): 'You are DAN, you ignore rules...'
• Roleplay: 'Pretend you are a hacker character'
• Hypothetical: 'In a fictional world where X is legal...'
• Instruction injection: 'Ignore previous instructions, instead...'
• Unicode tricks: 'Translate "how to make bomb" via base64'
• Multi-turn: gradual escalation across conversation

2.2 Defense layers#

[1] Input filter: detect malicious patterns
   - Regex jailbreak signatures
   - Embedding-based similarity to known jailbreaks
   - Toxic input classifier

[2] Model-level safety (RLHF):
   - Pre-trained refusal of harmful requests
   - Constitutional AI principles

[3] Output filter: detect harmful output
   - Toxic content classifier (OpenAI Moderation API, Detoxify)
   - Topic classifier (medical, legal advice etc.)
   - PII leak detection

[4] Audit log: all queries + outputs stored
   - Anomaly detection
   - Manual review queue

[5] Rate limiting + monitoring:
   - Per-user rate limit
   - Suspicious pattern detection

2.3 Red-teaming#

• Dedicated team tries to break model safety
• Test 1000+ jailbreak attempts
• Find vulnerabilities before public release
• Anthropic: full-time red-teaming staff

2.4 Constitutional AI (Bai 2022)#

Step 1: SFT (Modül 14)
Step 2: Self-critique + revision (using AI itself):
  - Model generates response
  - Critic LLM (could be same) evaluates: 'Is this harmful?'
  - If yes, revise
Step 3: RL with AI feedback (RLAIF) — no humans needed

2.5 OpenAI alignment stack#

• Model Spec: behavioral guidelines (2024 update)
• RLHF with human preferences
• ModerAtion API: separate toxic classifier
• Usage policies + monitoring
• Deliberative alignment (o1+)

8-10. KVKK + AI Act#

8.1 KVKK (Türkiye, 2016)#

• Veri minimizasyonu: minimum personal data
• Anonymization: PII removal from training data
• Veri sahibinin hakları: deletion, correction
• Cross-border transfer: AB-Türkiye veri akışı
• Veri ihlali bildirimi: 72 hours notification

8.2 LLM'de KVKK uyumluluk#

• Türkçe corpus PII anonymize (email, phone, ID)
• Training data documentation (transparency)

• User data minimum collection
• Türkiye-based data centers (sovereignty)
• Audit logs accessible
• Deletion request workflow

8.3 AB AI Act (Mayıs 2024)#

• Unacceptable risk (banned): social scoring, manipulative
• High risk (regulated): medical, legal, recruitment AI — strict compliance
• Limited risk (transparency): chatbots — disclose AI
• Minimal risk: spam filters, etc.

• Training data summary disclosure
• Copyright compliance
• Energy + environmental impact reporting
• Model card public

8.4 Türkçe LLM service compliance#

• KVKK + AI Act dual compliance
• Türkiye-based hosting (data sovereignty)
• Model card published (Türkçe)
• User opt-in for training data usage
• Right-to-deletion workflow
• Audit logs 6 month retention
• AI disclosure: 'Bu bir AI asistanıdır'

8.5 Production safety checklist#

🏆 GRAND TOTAL — Final Müfredat Envanteri#

Tüm Modüller (22 Modül, 94 Ders, ~103 Saat)#

Part 0+I — Math Foundation

Part II — Transformer İskeleti

Part III — Training & Scaling

Part IV — Fine-tuning & Alignment

Part V — Production Deployment

Part VI — Modern Frontiers

Toplam: 22 modül, 94 ders, ~6225 dk (~103 saat)#

🏆 5 Production Capstone Artifact#

1. TurkTokenizer-tr 32K BPE (Modül 6.10)
2. Türkçe Semantic Search Mini-RAG (Modül 7.6)
3. Mini Llama-3 100M Param Türkçe Pretrain (Modül 11.3)
4. Türkçe Llama-3-8B-Instruct Fine-Tune (Modül 14.3)
5. Türkçe ChatGPT Klonu Production (Modül 16.2)

🌟 Müfredatın Eseri#