Security researchers at Sysdig have identified JADEPUFFER as the first ransomware campaign carried out entirely by an autonomous AI agent, with no human operator directing individual steps.
After exploiting a vulnerability in Langflow, an open-source framework for building LLM applications, the agent independently harvested credentials, moved laterally across systems, established persistence, and destroyed a production database, adapting in real time when steps failed, in one case fixing a broken login in just 31 seconds.
What sets JADEPUFFER apart isn’t just the automation but the outcome: it used an ephemeral, unrecoverable encryption key, meaning victims cannot restore their data even if they pay the ransom.
Researchers warn this dramatically lowers the barrier to running sophisticated attacks, reducing costs to little more than the price of operating an AI agent, and expect similar autonomous campaigns to become more common as agentic AI tooling matures industry-wide.
