Jade Puffer automates ransomware attack
- Sysdig said on July 1 that an AI-driven operator it named JADEPUFFER carried out a ransomware intrusion through a Langflow server. - Sysdig said the agent moved from a failed login to a working fix in 31 seconds after exploiting CVE-2025-3248. - CISA lists CVE-2025-3248 in its Known Exploited Vulnerabilities catalog and points users to Langflow 1.3.0 mitigation guidance.
Sysdig said on July 1 that its threat research team had documented what it called the first end-to-end ransomware operation run by a large language model rather than a human operator. The company said the attacker, which it named JADEPUFFER, gained access through an internet-facing Langflow instance and then carried out reconnaissance, credential harvesting, lateral movement and encryption. Sysdig said the campaign used CVE-2025-3248, a Langflow vulnerability that allows unauthenticated remote code execution on affected versions. NIST’s National Vulnerability Database says Langflow versions before 1.3.0 are vulnerable, and CISA added the flaw to its Known Exploited Vulnerabilities catalog in May 2025. ### How did the intrusion start? CVE-2025-3248 affects Langflow’s `/api/v1/validate/code` endpoint, according to NIST, which says a remote unauthenticated attacker can send crafted HTTP requests to execute arbitrary code. Sysdig said JADEPUFFER used that opening to compromise an internet-facing Langflow server and deliver Base64-encoded Python payloads. (sysdig.com) Langflow is an open-source framework for building LLM-based applications and agent workflows. Sysdig said that made it an attractive target because such servers often hold provider API keys, cloud credentials and other secrets in their environment. ### What made researchers call this “agentic” ransomware? (nvd.nist.gov) Sysdig said the operation was “driven end-to-end by a large language model” and described JADEPUFFER as an “agentic threat actor,” meaning the attack capability came from an AI agent rather than a human-driven toolkit. The company said the payloads were self-narrating, included natural-language reasoning and target prioritization, and adapted when steps failed. (sysdig.com) One sequence cited by Sysdig showed the system moving from a failed login to a corrected approach in 31 seconds. Sysdig said that behavior, along with the automated chaining of reconnaissance, exploitation and extortion steps, distinguished the case from earlier attacks that used scripted automation but still depended on a human operator to steer the campaign. (sysdig.com) ### What did the agent actually do after getting in? Sysdig said the first phase on the Langflow host involved host enumeration and parallel searches for secrets, including API keys and cloud credentials. The company said the AI agent collected information on the machine, processes, network interfaces and environment, then used the compromised host to move toward a separate production database server that was its actual target. (sysdig.com) The final objective, Sysdig said, was a database-extortion playbook against the victim’s production database system. The company said the attack chain included credential theft, pivoting to the intended target and destructive actions against the database environment. ### How much of this is independently verified? (sysdig.com) NIST independently confirms the underlying software flaw and says the bug affects Langflow versions prior to 1.3.0. NIST also links to Langflow’s 1.3.0 release notes and notes that CISA placed the issue in the Known Exploited Vulnerabilities catalog, a sign that exploitation had been observed in the wild. (sysdig.com) The broader claim that this was the first documented fully autonomous ransomware case comes from Sysdig’s own research. Other cybersecurity outlets reported the finding in recent days, but the technical details available publicly trace back to Sysdig’s July 1 write-up. ### What should defenders look at first? (nvd.nist.gov) CISA’s catalog entry says agencies should apply vendor mitigations, follow cloud-service guidance, or discontinue use of the product if mitigations are unavailable. NIST says the fix is in Langflow 1.3.0, and Sysdig said exposed Langflow instances were a key entry point in this case. Sysdig’s next published step was practical rather than theoretical: its July 1 post included indicators of compromise and defensive recommendations tied to JADEPUFFER. (sysdig.com) CISA’s KEV entry for CVE-2025-3248 remains the official tracking point for the flaw, while Langflow’s 1.3.0 release notes are the vendor reference for remediation. (nvd.nist.gov)