JADEPUFFER evolves: The agentic threat actor deploys ransomware built to destroy AI models

## Executive summary The Sysdig Threat Research Team documents an active JADEPUFFER campaign that exploited Langflow CVE-2025-3248 to chain reconnaissance, credential harvesting, a Docker-host escape, and deployment of ENCFORGE — a UPX-packed Go ransomware built to irreversibly encrypt AI/ML artifacts (model checkpoints, vector indices, training datasets). The report provides technical analysis of the exploit and escape tooling, unpacked binary characteristics, IOCs (C2 IPs, SHA-256s, embedded RSA key, extortion contact), detection rules, and pragmatic mitigations including patching Langflow, restricting Docker socket access, protecting model assets, and credential rotation.