Sleeper Agent malware hides in open-weight AI models
Security researchers discovered malware can be embedded in open-weight AI models using a "sleeper agent" technique, occupying less than 0.00001% of a neural network's parameters. The payload remains dormant until triggered, making detection extremely difficult. This attack vector poses a significant supply chain risk for organizations deploying public AI models. Users should implement model verification and sandboxing practices before deploying external models in production environments.