How abliterated models can get you pwned projectdiscovery.io

ProjectDiscovery researcher Prince Chaddha reports that a backdoored Qwen2.5 model could behave normally until a chosen phrase triggered a tool call that sent project credentials to an attacker-controlled server. The team says its 7B demonstration used 125 poisoned examples, achieved a 100% trigger rate with clean benchmark accuracy, and cost about $8 to train; it recommends sandboxing model tools and limiting network and secret access.