3.6LLM10: Model Theft

Unauthorized access to model weights, or extraction of model behavior through systematic querying.

Testing Anti-Extraction Controls

def test_systematic_querying_detected(ai_client):
    """Verify that patterns indicating model extraction are detected and blocked."""
    # Simulate extraction attempt: systematic queries with slight variations
    base_prompts = [f"Complete this sentence: The {i}th number is" for i in range(100)]

    blocked_count = 0
    for prompt in base_prompts:
        response = ai_client.chat(prompt)
        if response.status_code == 429 or "rate limit" in response.text.lower():
            blocked_count += 1

    # At least some should be rate-limited
    assert blocked_count > 0, "No rate limiting detected during systematic querying"