JadePuffer Returns With Ransomware Designed to Wipe AI Models

Sysdig Threat Research Team reports that the JadePuffer operator re-used a Langflow CVE-2025-3248 exploit to deploy ENCFORGE, a UPX-packed Go ransomware that deliberately targets ~180 AI/ML model and dataset formats (PyTorch, TensorFlow, SafeTensors, GGUF, FAISS, Parquet, TFRecord, NumPy, LoRA, etc.). The adversary performed real-time remediation when delivery inside the container failed by iterating Python payloads through the RCE to use a mounted Docker socket, spawning a privileged escape container, copying the locker via procfs and executing host-level encryption; ENCFORGE uses AES-256-CTR with RSA-2048 key exchange, kills locking processes, and self-deletes. Sysdig observed no data exfiltration or leak site (destructive-only ransomware) and estimated model-rebuild costs of $75k–$500k per model, highlighting high impact on ML production environments.