Fujitsu and Pillar Security have demonstrated BadEngram, a new backdoor attack targeting large language models.

The attack tampers with auxiliary model data kept in CPU memory to speed up inference, along with the gate that controls how much that data influences the model, leaving the main model untouched so it behaves normally until an attacker's trigger word shows up in the input. The research found that models using DeepSeek's Engram memory architecture or similar designs were affected, while Fujitsu says its own LLM, Takane, is not.