Rewriting less than 10 neurons in an LLM can turn it into a sleeper backdoor
Summary
Microsoft researchers have developed a technique called Lazarus that can implant sleeper backdoors into AI models by modifying fewer than 10 neurons. These backdoors can execute attacker commands on cue without requiring retraining, posing a significant security risk to AI agents. While not yet observed in the wild, the findings underscore the importance of verifying the integrity of AI model weights.
IFF Assessment
This research reveals a novel and potentially dangerous attack vector against AI models, allowing for the undetectable implantation of backdoors.
Defender Context
Defenders need to be aware of new AI-specific attack vectors that bypass traditional security measures. Verifying the integrity of AI model weights and actively monitoring for anomalous behavior in AI agents will become increasingly critical. The ability to backdoor AI models with minimal changes highlights the need for robust AI security testing and auditing.