FuzzML

FuzzML is a continuous fuzzing project for the low-levels of machine learning security. We fuzz language models and looks both into the memory-layer and the semantic-layer, both traditional memory corruptions and model runtime behaviors.

FuzzML looks into:

• Language Model Format Parsing
• llama.cpp
• mlc-llm
• vllm
• Tokenizer Behaviors
• Adapters/LoRA Implementations