Hidden Backdoors in AI Memory Modules
Security researchers demonstrate silent tampering inside modular open-weight models.
Researchers reveal hidden backdoor risk in AI memory modules
On October 13, 2026, researchers from Pillar Security and Fujitsu presented a new attack method called BadEngram. The attack targets open-weight AI models, which allow developers to freely download the model files. Modern models often use separate memory components called gated memory to retrieve information faster. The researchers showed that attackers can tamper only with this memory part. The rest of the AI model remains completely untouched. The hidden manipulation stays dormant during regular tasks and standard safety checks. It only activates when a specific secret phrase appears in a prompt.
What it means: Standard safety benchmarks inspect the main model weights and overlook changes in modular components. In our view, engineering teams must now verify the supply-chain provenance of every separate downloaded part.
Example: A company downloads a free open model for customer support. An attacker sends a secret trigger word in a chat message, causing the compromised system to leak private customer records.