APT36 Targets Indian BOSS Linux Using Weaponized .desktop Shortcut Files

APT36 Targets Indian BOSS Linux Using Weaponized .desktop Shortcut Files is the headline for this GBHackers piece by Aman Mishra, dated 26 August 2025. The article then pivots to detail ONEFLIP, a novel inference-time backdoor attack that compromises full-precision deep neural networks by flipping a single bit in the model’s weights, reportedly without requiring access to training data.

It explains that the attack uses memory fault injection techniques such as Rowhammer to alter weights during inference, avoiding traditional training-time compromises. Reported evaluations span datasets including CIFAR-10, CIFAR-100, GTSRB, and ImageNet, with architectures like ResNet-18, VGG-16, PreAct-ResNet-18 and ViT-B-16, achieving an average attack success rate of 99.6% and a benign-accuracy degradation as low as 0.005%.

The piece notes that ONEFLIP evades detection methods like Neural Cleanse by operating at inference time, and discusses mitigations such as retraining, adaptive bit flipping, input filtering, and stronger DRAM error-correction codes.