You can capture and block backdoors like these and others with neuron activation-based signal monitoring at inference, with zero overhead. We have done this at WoolyAI. We built a poisoned LoRa adapter on Qwen, then caught it with our model signal monitoring. This monitoring wasn't purpose-built for this backdoor, but for generic unsafe/cyber signals.
You can capture and block backdoors like these and others with neuron activation-based signal monitoring at inference, with zero overhead. We have done this at WoolyAI. We built a poisoned LoRa adapter on Qwen, then caught it with our model signal monitoring. This monitoring wasn't purpose-built for this backdoor, but for generic unsafe/cyber signals.