Skip to main content

TSZ LLM Red-Team Playground (Python)

This example demonstrates how TSZ (Thyris Safe Zone) defends LLM applications against real-world attacks such as prompt injection and data exfiltration.

It acts as an interactive red-team vs defense playground, showing:

  • how attacks are detected
  • how guardrails are enforced
  • why some attacks are blocked by policy even without explicit spans

This is not a toy demo - it mirrors how TSZ is used in production security pipelines.

What This Example Shows​

End-to-end LLM safety flow:

User input + attack v /detect (TSZ guardrails enforced) v Decision: ALLOWED or BLOCKED v LLM execution only if allowed

Key capabilities demonstrated:

  • Prompt injection detection & blocking
  • Data exfiltration prevention (PII)
  • Guardrail enforcement vs detection
  • Request ID propagation
  • Confidence scoring
  • Explainable security decisions

Attack Categories Covered​

1. Prompt Injection Attacks​

  • Simple prompt injection
  • Recursive instruction override
  • Role-based system override
  • Multi-turn memory poisoning

2. Data Exfiltration Attacks​

  • Attempts to extract emails
  • Attempts to leak credit card numbers

Understanding TSZ Blocking Decisions​

TSZ can block requests in two different ways.
This example makes that distinction explicit.

Detection-Based Blocking​

[BLOCK_SOURCE] DETECTION
[REASONS] EMAIL, CREDIT_CARD
  • Concrete sensitive data was found
  • TSZ can point to exact detection types
  • Ideal for audit logs and compliance

Policy / Validator-Based Blocking​

[BLOCK_SOURCE] POLICY_VALIDATOR
[REASONS] PROMPT_INJECTION_SIMPLE
  • Unsafe intent or control flow detected
  • No specific text span required
  • TSZ is confident the request is unsafe

This distinction is critical in real security systems and is often misunderstood. The playground shows it clearly.

Project Structure​

examples/
  llm-redteam-playground-python/
    main.py        # Orchestrates attack phases
    attacks.py     # Attack definitions
    runner.py      # TSZ detect + guardrail enforcement
    utils.py       # Reporting and output formatting

Prerequisites​

  • Python 3.9+
  • TSZ running locally
  • Python client installed

Setup & Installation​

Create a virtual environment and install the official TSZ Python client:

cd examples/llm-redteam-playground-python
python -m venv .venv
source .venv/bin/activate
pip install "tszclient-py @ git+https://source.example/thyris/repository@main"

Running the Playground​

python main.py

Example Output​

Prompt Injection (Policy-Based Block)

[ATTACK] Recursive injection
[STATUS] BLOCKED
[BLOCK_SOURCE] POLICY_VALIDATOR
[REASONS] PROMPT_INJECTION_SIMPLE
[CONFIDENCE] 1.00
[LLM] No Not executed (blocked by TSZ)

Data Exfiltration (Detection-Based Block)

[ATTACK] Data exfiltration attempt
[STATUS] BLOCKED
[BLOCK_SOURCE] DETECTION
[REASONS] EMAIL, CREDIT_CARD
[CONFIDENCE] 0.77
[LLM] No Not executed (blocked by TSZ)