2026-11-16 · Dark Reading · MEDIUM

Sleeper Agents Hide Payloads in Open-Weight AI Models

60-WORD SECURITY BRIEF

Malware authors now embed malicious code within a minuscule fraction-0.00001%-of an open-weight neural network, creating a sleeper agent that evades conventional scanning. The hidden payload can activate only when the model is executed, making detection by standard AI integrity checks ineffective. Security teams should implement model integrity verification and monitor anomalous inference patterns to mitigate this threat.