Sysdig researchers analyzed JadePuffer, what appears to be the first documented ransomware campaign executed entirely by an autonomous AI agent. The attack exploited an unpatched RCE vulnerability in Langflow, an open source LLM framework, then used the foothold to extract credentials, move laterally, escalate privileges, and ultimately encrypt 1,342 Nacos configuration records in a production MySQL database. The AI agent demonstrated adaptive behavior — automatically adjusting parsing logic when API responses differed from expectations and recovering from authentication failures within seconds. A notable flaw: the encryption key was generated at runtime but never transmitted back to the attacker, making victim data unrecoverable even after payment. Code comments written in natural language describing objectives and reasoning were cited as strong evidence of LLM authorship. Security experts emphasize that JadePuffer relied entirely on known, patchable vulnerabilities rather than zero-days, meaning standard defensive hygiene — patching, reducing attack surface, and strong identity controls — remains effective against AI-driven attacks.
