Pillar Security Discovers ‘BadEngram,’ Exposing Hidden Supply-Chain Risk in Open-Weight AI Models

Researchers from Pillar Security and Fujitsu discovered BadEngram, a supply-chain attack that can implant hidden backdoors in open-weight AI models by modifying only memory components. The attack targets gated memory architectures used in models like Qwen3.8-Flash-Next and DeepSeek-V4.1-Flash, highlighting security risks in open-weight models as companies increasingly adopt them for cost savings.