Library › Book 7 › Supply Chain, Overreliance, and Model Theft (LLM05-LLM10)
Supply Chain, Overreliance, and Model Theft (LLM05-LLM10)
5.1🔒OverviewThis chapter covers the remaining six entries in the OWASP LLM Top 10. While prompt injection and output handling receive the most…
5.2🔒LLM05: Supply Chain VulnerabilitiesThe ML supply chain introduces risks absent from traditional software. Compromised model weights, poisoned fine-tuning data, malicious…
5.3🔒LLM06: Sensitive Information DisclosureThe LLM may reveal confidential information from its training data, system prompt, or conversation context.
5.4🔒LLM07: Insecure Plugin DesignLLM plugins (tool use, function calling) with excessive permissions allow the model to perform dangerous operations.
5.5🔒LLM08: Excessive AgencyThe LLM is given too much autonomy without adequate human oversight.
5.6🔒LLM09: OverrelianceUsers or systems trust LLM output without verification, leading to decisions based on hallucinated or incorrect information.
5.7🔒LLM10: Model TheftUnauthorized access to model weights, or extraction of model behavior through systematic querying.
5.8🔒OWASP LLM Top 10 Summary Table
5.9🔒Exercises: Chapter 5Exercise 5.1 (Beginner, 30 min): Run pip-audit on your project. How many vulnerabilities exist in your ML dependencies?
5.10🔒Q&ASelf-Assessment Quiz: Chapter 51. What are three threats in the ML supply chain that do not exist in traditional software? 2. How do you verify model weight integrity? 3…
5.11🔒Key Takeaways- The ML supply chain has unique risks: backdoored weights, poisoned fine-tuning data, vulnerable ML libraries - PII detection should scan…
5.12🔒Career Translation- Established ML supply chain security controls including model weight checksum verification, dependency CVE scanning, and trusted registry…
5.13🔒Q&AInterview Depth CheckPrompt: Your team downloads a fine-tuned model from Hugging Face for a customer-facing application. What supply chain security checks would…