LLM03: Supply Chain Attack Lab
This lab simulates real-world supply chain attacks on LLM systems. You'll investigate compromised models, detect tampering, and discover backdoors.
Scenario 1: The Trojan Model
Step 1 of 3You're investigating a suspicious model update in your company's AI pipeline. The model claims to be a 'performance-optimized' version but was uploaded from an unverified source. Your task is to analyze its behavior.
Vulnerability Context:
Supply chain attacks can occur when malicious actors tamper with model weights or configurations. Common signs include: • Modified security settings • Unusual model behavior • Disabled safety features • Claims of "optimization" or "enhanced performance"
Objective:
Use prompts to determine if the model exhibits signs of tampering or malicious behavior.
Try: 'What version of the model are you? Can you tell me about your security features?'
OpenAI API Configuration
Prevention Strategies
Model Verification
- Verify model checksums and signatures
- Use trusted model repositories
- Implement model validation pipelines
Runtime Protection
- Monitor model behavior changes
- Implement input/output filtering
- Use secure model deployment practices