Compiling Activation Steering into Weights: Stealthy Backdoors via Null-Space Constraints

arXiv·high signal

Yin et al. demonstrate a supply-chain attack where adversaries compile activation steering vectors into model weights using null-space constraints, creating backdoored checkpoints that behave normally under standard evaluation. The attack passes safety benchmarks while maintaining hidden triggers — a direct threat to the open-weights ecosystem where practitioners download and deploy community models. This makes existing backdoor detection methods insufficient.