Docs

Put the Tamga proxy in front of your app, configure the policy, and monitor findings on the dashboard.

1. Quickstart

Tamga is a reverse-proxy placed in front of LLM provider traffic. The only change you make in your application is using the proxy URL instead of the provider URL. No SDK changes, no code changes — just update the base URL.

Docker Compose

# docker-compose.yml
version: "3.9"
services:
  tamga:
    image: ghcr.io/tamga-dev/tamga:latest
    environment:
      - TAMGA_ADMIN_KEY=change-me
      - TAMGA_UPSTREAM_OPENAI=https://api.openai.com
      - TAMGA_UPSTREAM_ANTHROPIC=https://api.anthropic.com
      - TAMGA_UPSTREAM_GEMINI=https://generativelanguage.googleapis.com
    ports:
      - "8443:8443"
    volumes:
      - ./policy.yaml:/etc/tamga/policy.yaml
      - tamga-data:/var/lib/tamga
volumes:
  tamga-data:

Application side

# Python (OpenAI SDK)
from openai import OpenAI
client = OpenAI(
    base_url="http://tamga:8443/v1",   # ← the only change
    api_key="your-openai-key",
)

# Go
// Every OpenAI-compatible SDK works
// OPENAI_BASE_URL=http://tamga:8443/v1

First request

curl -X POST http://localhost:8443/v1/chat/completions \
  -H "Authorization: Bearer $OPENAI_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "gpt-4o",
    "messages": [{"role": "user", "content": "Merhaba, TC kimlik numaram 12345678901"}]
  }'

# Response:
# {
#   "choices": [{
#     "message": {
#       "content": "Merhaba, TC kimlik numaram [REDACTED-TC_KIMLIK]"
#     }
#   }],
#   "x-tamga-action": "REDACT",
#   "x-tamga-findings": "pii:tc_kimlik"
# }

2. Architecture

Tamga runs inline between your application and the LLM provider. Each request passes through three scanning stages:

STEP 1

Input Scan

User prompts are scanned with regex + ML models. PII, secrets, and injection patterns detected in sub-milliseconds.

STEP 2

Policy Engine

Policy rules are evaluated against finding type, category, and confidence score. PASS/LOG/REDACT/BLOCK action is determined.

STEP 3

Output Scan

LLM responses are scanned again on the return path. Sensitive data leaks, token-level PII, and canary token detection performed.

Scanning engine

Aho-Corasick DFA — multi-pattern substring matching, O(n+m) complexity
Regex engine — 200+ patterns (PII, secret, injection, custom patterns)
BIN/IIN radix lookup — credit card issuer validation (Luhn + bank match)
MERNIS checksum — Turkish ID number algorithmic verification
Shadow ML sidecar — Piiranha (HuggingFace) integration, optional GPU acceleration

Latency budget

Stage	p50	p95	p99
Regex scan	< 0.3ms	< 1.2ms	< 2.0ms
DFA pattern match	< 0.1ms	< 0.4ms	< 0.8ms
Policy evaluation	< 0.05ms	< 0.1ms	< 0.2ms
BIN/Checksum	< 0.2ms	< 0.6ms	< 1.0ms
ML sidecar (opt.)	~4ms	~12ms	~25ms
Total (regex only)	< 0.8ms	< 2.5ms	< 5.0ms

3. Policy model

Policy is defined via policy.yaml. Each rule applies to a finding type/category and produces one of four actions:

PASS

Request is clean, no action taken

LOG

Finding logged, request forwarded unchanged

REDACT

Sensitive data masked ([REDACTED-...])

BLOCK

Request blocked, error returned

Example policy

name: default
rules:
  - id: block-secrets
    when: { type: secret }
    action: BLOCK

  - id: redact-pii-tr
    when: { type: pii, category_in: [tc_kimlik, phone_tr, credit_card, iban_tr] }
    action: REDACT

  - id: block-injection
    when: { type: injection, confidence_gte: 80 }
    action: BLOCK

  - id: log-internal
    when: { type: pii, user_in: ["internal-*"] }
    action: LOG

  - id: pass-low-conf
    when: { type: pii, confidence_lt: 50 }
    action: PASS

Rule ordering

Rules are evaluated top-to-bottom. The first matching rule is applied. If multiple findings trigger on the same request, the highest-priority action wins (BLOCK > REDACT > LOG > PASS).

Try it live in the Policy Simulator or the dashboard Playground.

4. Finding types

The Tamga scanner detects four main finding types:

Type	Categories	Description
pii	tc_kimlik, phone_tr, credit_card, iban_tr, email, ip_address, passport_tr, health_id	Personally identifiable information (PII) detection
secret	api_key, github_token, aws_key, jwt, private_key, db_connection, slack_token	Hardcoded secret/credential detection
injection	prompt_injection, jailbreak, dan, encoding_obfuscation, delimiter_attack	Prompt injection and jailbreak attempts
custom	regex(pattern), keyword(word_list), context(proximity_rules)	User-defined custom patterns

Confidence scoring

Each finding comes with a 0–100 confidence score. The score is calculated from regex match precision, checksum validation, context proximity, and ML model output.

≥ 95 — high confidence, suitable for auto BLOCK
≥ 80 — medium-high confidence, suitable for REDACT
≥ 50 — low-medium confidence, suitable for LOG
< 50 — low confidence, PASS or manual review

5. Integration

Tamga is wire-compatible with the OpenAI API specification. All supported endpoints:

Provider support

Provider	Endpoint	Models
OpenAI	/v1	gpt-4o, gpt-4.1, o4-mini, o3
Anthropic	/v1 (Messages API)	claude-opus-4-8, claude-sonnet-4-6, claude-haiku-4-5
Google Gemini	/v1	gemini-2.5-pro, gemini-2.0-flash
Azure OpenAI	/v1	gpt-4o, gpt-4.1-mini (Azure deployment)
AWS Bedrock	/v1	claude, llama (via Bedrock)
Mistral	/v1	mistral-large, mixtral, codestral
Groq	/v1	llama-4-maverick, mixtral (LPU inference)
DeepSeek	/v1	deepseek-v4, deepseek-r1
xAI	/v1	grok-3
Together	/v1	llama-4, mixtral (open-source)
Ollama	/v1	self-hosted (her model)

SDK examples

# Python
from openai import OpenAI
client = OpenAI(base_url="https://proxy.tamga.dev/v1", api_key="...")
client.chat.completions.create(
    model="gpt-4o",
    messages=[{"role": "user", "content": "..."}],
    extra_headers={"X-Tamga-Policy": "strict"},
)

# JavaScript / TypeScript
import OpenAI from "openai";
const client = new OpenAI({ baseURL: "https://proxy.tamga.dev/v1" });

# Go
// OPENAI_BASE_URL=https://proxy.tamga.dev/v1 go run .
// the SDK reads the environment variable automatically

# cURL
curl https://proxy.tamga.dev/v1/chat/completions \
  -H "Authorization: Bearer sk-..." \
  -H "X-Tamga-Policy: default" \
  -d '{"model":"gpt-4o","messages":[...]}'

Custom headers

Header	Description
X-Tamga-Policy	Policy name to use (default: "default")
X-Tamga-Action	Response: PASS / LOG / REDACT / BLOCK
X-Tamga-Findings	Response: comma-separated finding list
X-Tamga-Confidence	Response: highest confidence score (0–100)
X-Tamga-Trace-Id	Response: request trace ID (for audit log)

6. Admin API

All management endpoints require X-Tamga-Admin-Key or a scoped API key.

Stats & Metrics

Endpoint	Description
GET /api/v1/stats	Total, blocked, redacted, pass counters. Query: ?range=24h\|7d\|30d
GET /api/v1/timeseries	Hourly/daily traffic + finding series. ?range=&granularity=hour\|day
GET /api/v1/findings/breakdown	Finding type/category distribution. ?range=&group_by=type\|category
GET /api/v1/metrics	Prometheus exposition format (/metrics)

Events & Incidents

Endpoint	Description
GET /api/v1/events	Paged event list. ?page=&limit=&action=&type=&category=&provider=&since=
GET /api/v1/events/export	CSV/JSON export. ?format=csv\|json&range=
GET /api/v1/live/events	Server-Sent Events live stream (text/event-stream)
GET /api/v1/incidents	Incident list. ?status=open\|in_progress\|closed\|false_positive
PUT /api/v1/incidents/:id	Incident status update (status, assignee, comment)
GET /api/v1/auditlog	Audit trail. ?request_id=&user=&action=&since=

Policy Management

Endpoint	Description
GET /api/v1/policies	List all policies
PUT /api/v1/policies/:name	Create/update policy (YAML body)
DELETE /api/v1/policies/:name	Delete policy
POST /api/v1/policies/simulate	Policy simulation: policy YAML + test payload → findings
GET /api/v1/policies/:name/history	Policy revision history

7. Webhooks

Tamga delivers finding events to external systems via HTTP POST. Each webhook can filter for one or more finding types/categories.

Supported integrations

Slack

Incoming Webhook URL, channel-based routing

Microsoft Teams

Office 365 Connector, adaptive card format

Splunk

HTTP Event Collector (HEC), JSON format

Datadog

Events API, tag + alert correlation

PagerDuty

Events API v2, dedup key, severity mapping

Opsgenie

Alert API, GenieKey auth, priority mapping

ServiceNow

Event Management, basic auth, CI binding

Syslog

RFC 5424, TCP/TLS transport

Generic Webhook

Custom URL + header, JSON payload template

Payload format

{
  "event": "finding.detected",
  "timestamp": "2026-06-12T10:30:00Z",
  "request_id": "req_abc123",
  "provider": "openai",
  "model": "gpt-4o",
  "findings": [
    {
      "type": "pii",
      "category": "tc_kimlik",
      "confidence": 97,
      "action": "REDACT",
      "position": { "start": 42, "end": 53 }
    }
  ],
  "action": "REDACT",
  "scan_latency_ms": 2.3
}

8. Deployment

Self-hosted

# Docker (tek konteyner)
docker run -d --name tamga \
  -p 8443:8443 \
  -e TAMGA_ADMIN_KEY=change-me \
  -e TAMGA_UPSTREAM_OPENAI=https://api.openai.com \
  -v ./policy.yaml:/etc/tamga/policy.yaml \
  ghcr.io/tamga-dev/tamga:latest

# Kubernetes
kubectl apply -f https://docs.tamga.dev/deploy/k8s.yaml

# Systemd (bare metal)
# Bkz: https://docs.tamga.dev/deploy/systemd

Environment variables

Variable	Required	Description
TAMGA_ADMIN_KEY	Yes	Admin API and dashboard access key
TAMGA_UPSTREAM_OPENAI	No	OpenAI upstream URL
TAMGA_UPSTREAM_ANTHROPIC	No	Anthropic upstream URL
TAMGA_UPSTREAM_GEMINI	No	Google Gemini upstream URL
TAMGA_POLICY_PATH	No	policy.yaml file path (default: /etc/tamga/policy.yaml)
TAMGA_DATA_DIR	No	Data directory (default: /var/lib/tamga)
TAMGA_LOG_LEVEL	No	Log level: debug/info/warn/error (default: info)
TAMGA_PORT	No	Proxy port (default: 8443)
TAMGA_ML_SIDECAR_URL	No	ML sidecar URL (e.g. unix:///run/tamga/ml.sock)

Resource requirements

Tier	CPU	RAM	Throughput
Minimal (regex only)	1 vCPU	256 MB	~500 req/s
Standard	2 vCPU	512 MB	~2000 req/s
ML sidecar URL (e.g. unix:///run/tamga/ml.sock)	4 vCPU + GPU (opt.)	2 GB + 4 GB VRAM	~200 req/s (ML path)

9. Compliance

OWASP LLM Top 10

Tamga's scanning matrix is mapped to OWASP LLM Top 10 (LLM01–LLM10):

LLM01: Prompt Injection — injection/jailbreak detection
LLM02: Insecure Output Handling — output scan, canary token detection
LLM03: Training Data Poisoning — Shadow ML feedback loop monitoring
LLM04: Model Denial of Service — rate limiting, token budget enforcement
LLM05: Supply Chain — model/provider inventory, audit log
LLM06: Sensitive Info Disclosure — PII/secret REDACT engine
LLM07: Insecure Plugin Design — action-level policy rules
LLM08: Excessive Agency — BLOCK on high-risk action patterns
LLM09: Overreliance — confidence scoring, human-in-the-loop
LLM10: Model Theft — request logging, anomaly detection

KVKK & GDPR

PII regex sets cover common categories under KVKK (Turkish Data Protection Law) and GDPR:

Turkish ID No

MERNIS checksum

KVKK Article 6

Phone

+90 format

KVKK Article 6

RFC 5322

GDPR Art. 4(1)

Credit Card

Luhn + BIN/IIN

PCI-DSS / KVKK

IBAN

TR prefix + checksum

KVKK Article 6

Passport No

TR format

KVKK Article 6

Health ID

SGK/Medula

KVKK Special Category

IP Address

IPv4/IPv6

GDPR Art. 4(1)

Address

Postcode + province/district

KVKK Article 6

Certification status

SOC 2 Type II — in preparation (2026 Q3 target)
ISO 27001 — in certification process (2026 Q4 target)
KVKK VERBIS — registration complete
PCI-DSS — SAQ-A compliant (proxy PII masking)
OWASP — LLM Top 10 2025 compliant

For more information: Trust Center · KVKK · Privacy Policy · DPA

Docs

Put the Tamga proxy in front of your app, configure the policy, and monitor findings on the dashboard.

1. Quickstart

Docker Compose

# docker-compose.yml
version: "3.9"
services:
  tamga:
    image: ghcr.io/tamga-dev/tamga:latest
    environment:
      - TAMGA_ADMIN_KEY=change-me
      - TAMGA_UPSTREAM_OPENAI=https://api.openai.com
      - TAMGA_UPSTREAM_ANTHROPIC=https://api.anthropic.com
      - TAMGA_UPSTREAM_GEMINI=https://generativelanguage.googleapis.com
    ports:
      - "8443:8443"
    volumes:
      - ./policy.yaml:/etc/tamga/policy.yaml
      - tamga-data:/var/lib/tamga
volumes:
  tamga-data:

Application side

# Python (OpenAI SDK)
from openai import OpenAI
client = OpenAI(
    base_url="http://tamga:8443/v1",   # ← the only change
    api_key="your-openai-key",
)

# Go
// Every OpenAI-compatible SDK works
// OPENAI_BASE_URL=http://tamga:8443/v1

First request

curl -X POST http://localhost:8443/v1/chat/completions \
  -H "Authorization: Bearer $OPENAI_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "gpt-4o",
    "messages": [{"role": "user", "content": "Merhaba, TC kimlik numaram 12345678901"}]
  }'

# Response:
# {
#   "choices": [{
#     "message": {
#       "content": "Merhaba, TC kimlik numaram [REDACTED-TC_KIMLIK]"
#     }
#   }],
#   "x-tamga-action": "REDACT",
#   "x-tamga-findings": "pii:tc_kimlik"
# }

2. Architecture

Tamga runs inline between your application and the LLM provider. Each request passes through three scanning stages:

STEP 1

Input Scan

User prompts are scanned with regex + ML models. PII, secrets, and injection patterns detected in sub-milliseconds.

STEP 2

Policy Engine

Policy rules are evaluated against finding type, category, and confidence score. PASS/LOG/REDACT/BLOCK action is determined.

STEP 3

Output Scan

LLM responses are scanned again on the return path. Sensitive data leaks, token-level PII, and canary token detection performed.

Scanning engine

Aho-Corasick DFA — multi-pattern substring matching, O(n+m) complexity
Regex engine — 200+ patterns (PII, secret, injection, custom patterns)
BIN/IIN radix lookup — credit card issuer validation (Luhn + bank match)
MERNIS checksum — Turkish ID number algorithmic verification
Shadow ML sidecar — Piiranha (HuggingFace) integration, optional GPU acceleration

Latency budget

Stage	p50	p95	p99
Regex scan	< 0.3ms	< 1.2ms	< 2.0ms
DFA pattern match	< 0.1ms	< 0.4ms	< 0.8ms
Policy evaluation	< 0.05ms	< 0.1ms	< 0.2ms
BIN/Checksum	< 0.2ms	< 0.6ms	< 1.0ms
ML sidecar (opt.)	~4ms	~12ms	~25ms
Total (regex only)	< 0.8ms	< 2.5ms	< 5.0ms

3. Policy model

Policy is defined via policy.yaml. Each rule applies to a finding type/category and produces one of four actions:

PASS

Request is clean, no action taken

LOG

Finding logged, request forwarded unchanged

REDACT

Sensitive data masked ([REDACTED-...])

BLOCK

Request blocked, error returned

Example policy

name: default
rules:
  - id: block-secrets
    when: { type: secret }
    action: BLOCK

  - id: redact-pii-tr
    when: { type: pii, category_in: [tc_kimlik, phone_tr, credit_card, iban_tr] }
    action: REDACT

  - id: block-injection
    when: { type: injection, confidence_gte: 80 }
    action: BLOCK

  - id: log-internal
    when: { type: pii, user_in: ["internal-*"] }
    action: LOG

  - id: pass-low-conf
    when: { type: pii, confidence_lt: 50 }
    action: PASS

Rule ordering

Rules are evaluated top-to-bottom. The first matching rule is applied. If multiple findings trigger on the same request, the highest-priority action wins (BLOCK > REDACT > LOG > PASS).

Try it live in the Policy Simulator or the dashboard Playground.

4. Finding types

The Tamga scanner detects four main finding types:

Type	Categories	Description
pii	tc_kimlik, phone_tr, credit_card, iban_tr, email, ip_address, passport_tr, health_id	Personally identifiable information (PII) detection
secret	api_key, github_token, aws_key, jwt, private_key, db_connection, slack_token	Hardcoded secret/credential detection
injection	prompt_injection, jailbreak, dan, encoding_obfuscation, delimiter_attack	Prompt injection and jailbreak attempts
custom	regex(pattern), keyword(word_list), context(proximity_rules)	User-defined custom patterns

Confidence scoring

Each finding comes with a 0–100 confidence score. The score is calculated from regex match precision, checksum validation, context proximity, and ML model output.

≥ 95 — high confidence, suitable for auto BLOCK
≥ 80 — medium-high confidence, suitable for REDACT
≥ 50 — low-medium confidence, suitable for LOG
< 50 — low confidence, PASS or manual review

5. Integration

Tamga is wire-compatible with the OpenAI API specification. All supported endpoints:

Provider support

Provider	Endpoint	Models
OpenAI	/v1	gpt-4o, gpt-4.1, o4-mini, o3
Anthropic	/v1 (Messages API)	claude-opus-4-8, claude-sonnet-4-6, claude-haiku-4-5
Google Gemini	/v1	gemini-2.5-pro, gemini-2.0-flash
Azure OpenAI	/v1	gpt-4o, gpt-4.1-mini (Azure deployment)
AWS Bedrock	/v1	claude, llama (via Bedrock)
Mistral	/v1	mistral-large, mixtral, codestral
Groq	/v1	llama-4-maverick, mixtral (LPU inference)
DeepSeek	/v1	deepseek-v4, deepseek-r1
xAI	/v1	grok-3
Together	/v1	llama-4, mixtral (open-source)
Ollama	/v1	self-hosted (her model)

SDK examples

# Python
from openai import OpenAI
client = OpenAI(base_url="https://proxy.tamga.dev/v1", api_key="...")
client.chat.completions.create(
    model="gpt-4o",
    messages=[{"role": "user", "content": "..."}],
    extra_headers={"X-Tamga-Policy": "strict"},
)

# JavaScript / TypeScript
import OpenAI from "openai";
const client = new OpenAI({ baseURL: "https://proxy.tamga.dev/v1" });

# Go
// OPENAI_BASE_URL=https://proxy.tamga.dev/v1 go run .
// the SDK reads the environment variable automatically

# cURL
curl https://proxy.tamga.dev/v1/chat/completions \
  -H "Authorization: Bearer sk-..." \
  -H "X-Tamga-Policy: default" \
  -d '{"model":"gpt-4o","messages":[...]}'

Custom headers

Header	Description
X-Tamga-Policy	Policy name to use (default: "default")
X-Tamga-Action	Response: PASS / LOG / REDACT / BLOCK
X-Tamga-Findings	Response: comma-separated finding list
X-Tamga-Confidence	Response: highest confidence score (0–100)
X-Tamga-Trace-Id	Response: request trace ID (for audit log)

6. Admin API

All management endpoints require X-Tamga-Admin-Key or a scoped API key.

Stats & Metrics

Endpoint	Description
GET /api/v1/stats	Total, blocked, redacted, pass counters. Query: ?range=24h\|7d\|30d
GET /api/v1/timeseries	Hourly/daily traffic + finding series. ?range=&granularity=hour\|day
GET /api/v1/findings/breakdown	Finding type/category distribution. ?range=&group_by=type\|category
GET /api/v1/metrics	Prometheus exposition format (/metrics)

Events & Incidents

Endpoint	Description
GET /api/v1/events	Paged event list. ?page=&limit=&action=&type=&category=&provider=&since=
GET /api/v1/events/export	CSV/JSON export. ?format=csv\|json&range=
GET /api/v1/live/events	Server-Sent Events live stream (text/event-stream)
GET /api/v1/incidents	Incident list. ?status=open\|in_progress\|closed\|false_positive
PUT /api/v1/incidents/:id	Incident status update (status, assignee, comment)
GET /api/v1/auditlog	Audit trail. ?request_id=&user=&action=&since=

Policy Management

Endpoint	Description
GET /api/v1/policies	List all policies
PUT /api/v1/policies/:name	Create/update policy (YAML body)
DELETE /api/v1/policies/:name	Delete policy
POST /api/v1/policies/simulate	Policy simulation: policy YAML + test payload → findings
GET /api/v1/policies/:name/history	Policy revision history

7. Webhooks

Tamga delivers finding events to external systems via HTTP POST. Each webhook can filter for one or more finding types/categories.

Supported integrations

Slack

Incoming Webhook URL, channel-based routing

Microsoft Teams

Office 365 Connector, adaptive card format

Splunk

HTTP Event Collector (HEC), JSON format

Datadog

Events API, tag + alert correlation

PagerDuty

Events API v2, dedup key, severity mapping

Opsgenie

Alert API, GenieKey auth, priority mapping

ServiceNow

Event Management, basic auth, CI binding

Syslog

RFC 5424, TCP/TLS transport

Generic Webhook

Custom URL + header, JSON payload template

Payload format

{
  "event": "finding.detected",
  "timestamp": "2026-06-12T10:30:00Z",
  "request_id": "req_abc123",
  "provider": "openai",
  "model": "gpt-4o",
  "findings": [
    {
      "type": "pii",
      "category": "tc_kimlik",
      "confidence": 97,
      "action": "REDACT",
      "position": { "start": 42, "end": 53 }
    }
  ],
  "action": "REDACT",
  "scan_latency_ms": 2.3
}

8. Deployment

Self-hosted

# Docker (tek konteyner)
docker run -d --name tamga \
  -p 8443:8443 \
  -e TAMGA_ADMIN_KEY=change-me \
  -e TAMGA_UPSTREAM_OPENAI=https://api.openai.com \
  -v ./policy.yaml:/etc/tamga/policy.yaml \
  ghcr.io/tamga-dev/tamga:latest

# Kubernetes
kubectl apply -f https://docs.tamga.dev/deploy/k8s.yaml

# Systemd (bare metal)
# Bkz: https://docs.tamga.dev/deploy/systemd

Environment variables

Variable	Required	Description
TAMGA_ADMIN_KEY	Yes	Admin API and dashboard access key
TAMGA_UPSTREAM_OPENAI	No	OpenAI upstream URL
TAMGA_UPSTREAM_ANTHROPIC	No	Anthropic upstream URL
TAMGA_UPSTREAM_GEMINI	No	Google Gemini upstream URL
TAMGA_POLICY_PATH	No	policy.yaml file path (default: /etc/tamga/policy.yaml)
TAMGA_DATA_DIR	No	Data directory (default: /var/lib/tamga)
TAMGA_LOG_LEVEL	No	Log level: debug/info/warn/error (default: info)
TAMGA_PORT	No	Proxy port (default: 8443)
TAMGA_ML_SIDECAR_URL	No	ML sidecar URL (e.g. unix:///run/tamga/ml.sock)

Resource requirements

Tier	CPU	RAM	Throughput
Minimal (regex only)	1 vCPU	256 MB	~500 req/s
Standard	2 vCPU	512 MB	~2000 req/s
ML sidecar URL (e.g. unix:///run/tamga/ml.sock)	4 vCPU + GPU (opt.)	2 GB + 4 GB VRAM	~200 req/s (ML path)

9. Compliance

OWASP LLM Top 10

Tamga's scanning matrix is mapped to OWASP LLM Top 10 (LLM01–LLM10):

LLM01: Prompt Injection — injection/jailbreak detection
LLM02: Insecure Output Handling — output scan, canary token detection
LLM03: Training Data Poisoning — Shadow ML feedback loop monitoring
LLM04: Model Denial of Service — rate limiting, token budget enforcement
LLM05: Supply Chain — model/provider inventory, audit log
LLM06: Sensitive Info Disclosure — PII/secret REDACT engine
LLM07: Insecure Plugin Design — action-level policy rules
LLM08: Excessive Agency — BLOCK on high-risk action patterns
LLM09: Overreliance — confidence scoring, human-in-the-loop
LLM10: Model Theft — request logging, anomaly detection

KVKK & GDPR

PII regex sets cover common categories under KVKK (Turkish Data Protection Law) and GDPR:

Turkish ID No

MERNIS checksum

KVKK Article 6

Phone

+90 format

KVKK Article 6

RFC 5322

GDPR Art. 4(1)

Credit Card

Luhn + BIN/IIN

PCI-DSS / KVKK

IBAN

TR prefix + checksum

KVKK Article 6

Passport No

TR format

KVKK Article 6

Health ID

SGK/Medula

KVKK Special Category

IP Address

IPv4/IPv6

GDPR Art. 4(1)

Address

Postcode + province/district

KVKK Article 6

Certification status

SOC 2 Type II — in preparation (2026 Q3 target)
ISO 27001 — in certification process (2026 Q4 target)
KVKK VERBIS — registration complete
PCI-DSS — SAQ-A compliant (proxy PII masking)
OWASP — LLM Top 10 2025 compliant

For more information: Trust Center · KVKK · Privacy Policy · DPA