safegate 1.0.0

Self-hosted security layer for LLM applications — detect PII by format, substitute or mask, and audit every call.

English | Русский

SafeGate

Self-hosted security layer for LLM applications

Detect PII by format, not labels · Substitute (not only ***) · RU + EN · Block leaks · Audit everything

Quick start · Examples · Compare · Benchmarks · Community

_{pip install safegate · Branding & i18n}

---

What it does

SafeGate sits between your app and the LLM:

Your text  →  detect PII  →  apply policy  →  safe prompt  →  LLM  →  check response  →  safe output

• Self-hosted — no cloud lock-in for sensitive data

• Substitution by default — realistic fakes so the model keeps structure (emails, IDs, cards)

• Format-based detection — catches values without labels like phone: or СНИЛС:

• RU + US + universal — INN, SNILS, SSN, IBAN, addresses, cards, and 70+ entity types

---

Live demo UI

Pipeline overview

Detections & substitution

The Pipeline Explorer (local, optional) shows all 8 stages: detectors → policy → prompt guard → mock LLM → response guard.

| Mode | Behavior |

|------|----------|

| Substitute | Replace PII with dictionary fakes (LLM-friendly) |

| Mask | Replace with *** |

python -m pip install -e .

cd local-demo && pip install -r requirements.txt && python app.py

# → http://127.0.0.1:8765

Regenerate README screenshots: python scripts/generate_demo_screenshots.py

---

Quick start

Install

pip install safegate

# or from source:

git clone https://github.com/safegate-ai/safegate.git && cd safegate

python -m pip install -e ".[dev]"

See docs/INSTALL.md · Release guide for PyPI publish.

Minimal example

from safegate import SafeGate

from safegate.llm import MockLLM



guard = SafeGate(protection_mode="substitute")  # or "mask"



response = guard.chat(

    "Reach me at [e-mail] or [phone]",

    MockLLM(),

)

print(response)

Fluent API

session = guard.protect_prompt("Card 4532 1987 4421 8426, exp 11/30")

print(session.prompt_mappings)

result = session.invoke(MockLLM()).protect_response()

Enterprise presets (optional)

guard = SafeGate(preset="health_hipaa", region="us")

---

Examples

| Example | Use case |

|---------|----------|

| chatbot_demo.py | Chatbot with PII substitution |

| rag_demo.py | RAG context guard |

| mcp_demo.py | MCP tool filter |

| gateway_client.py | REST AI gateway |

| langchain_demo.py | LangChain handler |

Full list: examples/README.md

---

How it works

flowchart LR

    A[Input text] --> B[Detectors]

    B --> C[Policy engine]

    C --> D{Action}

    D -->|substitute| E[Dictionary vault]

    D -->|mask| F[Asterisks]

    D -->|block| G[Stop]

    E --> H[LLM]

    F --> H

    H --> I[Response guard]

    I --> J[Output]

| Stage | Role |

|-------|------|

| Detectors | 70+ entity types — email, phone, INN/SNILS, SSN, IBAN, address, card, SWIFT, MAC/IP, names, … |

| Policy | substitute · mask · block · remove |

| Prompt guard | Before the LLM call |

| Response guard | Leak detection, detokenize, optional regenerate |

| Audit | JSONL, CEF / ECS / CSV export |

---

Compare with alternatives

| | DIY regex | Presidio | LLM Guard | SafeGate |

|---|:---:|:---:|:---:|:---:|

| Self-hosted | ✅ | ✅ | ✅ | ✅ |

| LLM session API | DIY | DIY | Partial | ✅ |

| Substitution (not only ***) | ❌ | Partial | ❌ | ✅ |

| RU + US regional IDs | Manual | Custom | Limited | ✅ |

| Response leak + retry | ❌ | Manual | ✅ | ✅ |

| YAML policies + simulator | ❌ | ❌ | ❌ | ✅ |

| Multi-tenant gateway + SIEM | ❌ | ❌ | ❌ | ✅ |

Full comparison: docs/COMPARISON.md · Benchmarks: docs/BENCHMARK.md

---

Industry packs

bank_ru · health_hipaa · gov_fz152 · retail_eu · fintech_us · insurance_us · saas_global · telecom_eu · education_us · energy_eu · legal_eu · logistics_eu

safegate policy catalog --format markdown

Entity catalog

---

Development

python -m pip install -e ".[dev]"

pytest -v

python scripts/benchmark.py

CONTRIBUTING.md · Good first issues · SECURITY.md · Community

---

License

MIT

---

LLM power without leaking PII.