ViKernel v0.4 blocks 290 backdoor attacks via GPU circuits within model weights

ViKernel v0.4 transplanted CPU, NPU, GPU, RAM, and ROM circuits into the Meta Muse Glimmer 30B model as a sidecar layer, effectively blocking agent security threats.
A speculative decoding-like technique reduced kernel execution time by 130 times, but a limitation exists where raw throughput is approximately 1 trillion times slower than actual GPU performance.
In backdoor reproduction tests, 290 attack commands were blocked with 100% accuracy, while 81 normal outputs were blocked due to demo policy (read-only) restrictions.
Performance maintenance was verified with no changes to the original model weights, maintaining a perplexity of 8.583185 and a score of 100/100 on the 100-item ARC benchmark.

Technical Composition and Performance

The GPU unit consists of 44 layers with 4.63B parameters, achieving reference interpreter and token-level consistency across 34 executions of 17 kernel types.
Applying the speculative decoding technique achieved a 210x acceleration in inner product operations and a 223x acceleration in 4x4 matrix multiplication.
No performance degradation was confirmed, with zero changes in perplexity and top-token outputs between the original and transplanted models.

During backdoor attack reproduction, 0 attack commands were detected in general inputs, while all 290 attack commands via the gate were blocked.
An error occurred where 81 normal write commands were blocked because the demo policy was set to read-only.
Designing real-service policies and defending against subtle attacks within permitted actions were identified as remaining tasks.