Quantization and Compression Attacks: How Model Size Reduction Can Re-Enable Suppressed Unsafe Behaviors
When you deploy a GGUF Q4_K_M quantization of a safety-aligned model, you're relying on an assumption: that alignment survives precision reduction. Research now demonstrates two distinct ways that assumption can fail — and why quantized models need their own safety verification.