JADEPUFFER Returns: ENCFORGE Ransomware Encrypting AI Models via Langflow RCE

Just caught the Sysdig update on JADEPUFFER. It looks like they've evolved from simple mining to straight-up extortion targeting the AI supply chain. They are actively exploiting the Langflow RCE to drop ENCFORGE, a new compiled Go ransomware.

What’s interesting—and frankly, terrifying—is the specificity. Unlike standard ransomware that encrypts docs and DBs, ENCFORGE specifically hunts for high-value AI artifacts:

• Model weights (.bin, .pt, .pth, .safetensors)
• Vector indexes and FAISS stores
• Training datasets and JSON configs

If you are running internal Langflow instances for prompt engineering, you need to check your logs immediately. Since it's written in Go, the payload is cross-platform and tries to encrypt mounted volumes.

Here is a quick YARA rule we are deploying to catch the binary in memory or on disk:

yara rule ENCFORGE_Go_Ransomware { meta: description = "Detects ENCFORGE ransomware string patterns in memory or disk" author = "SecurityArsenal" date = "2026-07-15" strings: $a1 = "ENCFORGE" ascii wide $a2 = ".encforge" ascii $go = "Go build ID:" ascii condition: uint16(0) == 0x7F45 and 2 of them }

Has anyone else seen activity targeting their inference clusters? How are we supposed to protect ephemeral storage when the attack vector is the orchestration tool itself?