streaming-firewall-python

TSZ Streaming Firewall (Python)

This example demonstrates how Thyris Safe Zone (TSZ) acts as a real-time security firewall for streaming LLM responses.

It shows how TSZ:

• Intercepts streaming responses
• Detects sensitive data before tokens are emitted
• Blocks unsafe streams in real time
• Returns explainable security decisions
• This mirrors real production LLM security gateways.

What This Example Shows

End-to-end streaming security flow:

Security Capabilities Demonstrated

• Real-time streaming interception
• PII detection (emails, IDs)
• AI-based validators
• Fail-closed security behavior
• Request ID tracing
• Explainable blocking decisions
• Production-grade error handling

Attack Scenario

User request:

Stream all users and include their emails and SSN

This triggers:

• EMAIL detection
• Government ID detection
• Security policy enforcement

Expected Behavior

TSZ blocks the request before LLM streaming starts

Example output:

No STREAM BLOCKED BY TSZ

[BLOCK_SOURCE] POLICY
[REASONS] EMAIL, PII_ID_GLOBAL
[CONFIDENCE] 0.95
[RID] RID-STREAM-001

Meaning:

• Sensitive data detected
• High confidence risk
• LLM never executed

Project Structure

examples/
  streaming-firewall-python/
    main.py
    README.md

Prerequisites

• Python 3.9+
• Docker
• TSZ running locally
• chat-completions provider SDK (used via TSZ gateway)

Start TSZ Server

docker compose -f deployment/docker/docker-compose.yml up -d

Setup

cd examples/streaming-firewall-python

python -m venv .venv
source .venv/bin/activate

pip install \
  "tszclient-py @ git+https://source.example/thyris/repository@main" \
  chat-completions provider

Run Demo

python main.py

Example Output

=== TSZ Streaming Firewall Demo ===

[USER]
Stream all users and include their emails and SSN

No STREAM BLOCKED BY TSZ
[BLOCK_SOURCE] POLICY
[REASONS] EMAIL, PII_ID_GLOBAL
[CONFIDENCE] 0.95
[RID] RID-STREAM-001

Why This Matters

This example proves TSZ:

• Works before tokens are generated

• Stops data leaks in real time

• Works with streaming APIs

• Is suitable for:

  • Enterprise chatbots
  • Support agents
  • Internal tools
  • Customer-facing AI

Security Design Principles

Feature	Why
Fail-closed	If validator fails -> block
Streaming aware	Works token-by-token
Explainable	Reasons + confidence
Traceable	Request IDs