Pillar Security Discovers 'BadEngram,' Exposing Hidden Supply-Chain Risk in Open-Weight AI Models – IT Business Net

The BadEngram attack demonstrates a critical vulnerability in emerging AI memory architectures, specifically targeting the gated memory components found in recent Qwen and DeepSeek models. By modifying only these specialized memory parameters, attackers can implant a persistent, hard-to-detect backdoor that remains dormant until a specific trigger is activated. This reveals that even minimal changes to non-core model structures can compromise system integrity, leaving the rest of the model seemingly unaffected while subtly altering its behavior. This discovery is particularly significant for the open_data community because it highlights the hidden security risks associated with open-weight models. As organizations increasingly adopt these models for cost efficiency and transparency, they assume full responsibility for validating their safety. The attack exposes a new attack surface where novel architectural components lack the rigorous security scrutiny given to traditional model backbones, emphasizing that open access does not equate to inherent security. Consequently, the article underscores the urgent need to expand threat modeling to include the entire AI system, encompassing memory and software supply chains. Security controls must verify the integrity and provenance of all learned components, especially when models are modified or sourced from third parties. This serves as a vital warning for open-data practitioners to treat complete models as supply-chain artifacts, ensuring that the move toward open-weight AI does not inadvertently introduce systemic vulnerabilities.

Source: itbusinessnet.com
Published on 2026-10-15

Pillar Security Discovers 'BadEngram,' Exposing Hidden Supply-Chain Risk in Open-Weight AI Models – IT Business Net

Related news