ECLYPSIUM SHOWS AI MODEL WEIGHTS CAN BE ALTERED IN MEMORY AFTER INTEGRITY CHECKS PASS

Summary from the TLDR Hardware newsletter

Eclypsium demonstrated changing a model's weights in memory after integrity checks pass, so the model file, container, hash, and signatures on disk all stay valid. Attack paths run through host firmware, management controllers, and DPUs, and the report says Nvidia, AMD, Intel, and Arm systems are all exposed. Its fix is DPUs that enforce security policy independently of the AI system, plus full component inventories.