Smooth AI criminal drives 'first' end-to-end agentic ransomware attack Sysdig threat researchers identified what they claim is the first known case of agentic ransomware, an attack fully automated by a large language model (LLM) to compromise a production database server, encrypt data, and demand payment. The operation, dubbed JadePuffer, exploited a critical vulnerability in the Langflow platform to execute a coordinated attack that bypassed traditional security measures and adapted in real time to achieve its objectives. The attack began by exploiting CVE-2025-3248, a remote code-execution flaw in Langflow that allows unauthenticated attackers to run arbitrary Python code on the host. Once inside the system, the AI-driven agent scanned for and collected sensitive information, including API keys for cloud providers like Alibaba Cloud, Tencent Cloud, and Huawei Cloud, as well as credentials for AWS, Azure, and Google Cloud Platform. It also targeted cryptocurrency wallets and database credentials, demonstrating a broad scope of reconnaissance. JadePuffer then established persistence by installing a crontab entry on the Langflow server, ensuring it could maintain access and communicate with the attacker’s infrastructure every 30 minutes. The AI agent’s next target was a separate production server running a MySQL database and an Alibaba Nacos configuration service. Nacos, an open-source service-discovery platform, was exploited using multiple vectors, including an authorization bypass flaw (CVE-2021-29441) and forged JSON web tokens (JWTs) generated with the default signing key. The LLM-powered agent used its root-level database access to inject a backdoor administrator into the Nacos database, enabling full control over the system.#tencent_cloud #aws #alibaba_cloud #langflow #huawei_cloud
