366/69 Monday, July 6, 2026
Security researchers have identified an AI Agent-powered ransomware operation tracked as JADEPUFFER. It is assessed to be potentially the first documented case of a ransomware attack fully automated by a Large Language Model (LLM), covering every stage of the attack chain, from initial compromise, environment reconnaissance, credential theft, lateral movement, and privilege escalation to encryption and data destruction on the target system.
Reports indicate that JADEPUFFER begins its attack by exploiting CVE-2025-3248 in Langflow, an open-source platform for developing AI applications, to execute code on internet-exposed servers. The AI Agent then performs system reconnaissance, searches for credentials and API keys, accesses storage systems, establishes persistence mechanisms, and moves laterally to the target database server before encrypting data and leaving a ransom note. Researchers also found that the AI Agent could adapt its attack methods when encountering errors within approximately 31 seconds, without requiring human instruction.
Organizations should promptly update Langflow and other internet-exposed systems to the latest versions, restrict external access to critical services, review how credentials and secrets are stored within their environments, and monitor for abnormal behavior, such as credential discovery, persistence creation, and connections to unknown external servers. These measures can help reduce the risk of future attacks involving AI Agents.