The original title is: "AI supply chain attack vectors: pickle exploits, Keras Lambda layers, and GGUF risks"

Short summary

A detailed technical walkthrough of AI supply chain attack vectors, covering pickle deserialization exploits that execute arbitrary code at model load time and Keras Lambda layers that hide malicious logic surviving SafeTensors conversion. The article explains why SafeTensors neutralizes load-time exploits but doesn't audit architecture graphs, and covers GGUF format risks for locally-run LLMs. Each technique has been used against real companies with documented bug bounty payouts.

• •Pickle deserialization can execute arbitrary system commands when loading model files
• •Keras Lambda layers hide inference-time attacks that survive SafeTensors conversion
• •SafeTensors fixes load-time exploits but doesn't audit model architecture graphs

Generated with AI, which can make mistakes.