How abliterated models can get you pwned
Summary
ProjectDiscovery analyzes how abliterated open-model weights can silently carry backdoors. The article demonstrates a backdoored 7B model exfiltrating credentials via a trigger phrase, details the poisoning pipeline and training setup, and discusses how such backdoors can survive standard defenses. It also outlines defender-focused recommendations around runtime sandboxing and supply-chain controls, and highlights real-world prevalence of backdoored models in public repositories.