The first documented case of an end-to-end ransomware operation executed autonomously by a large language model (LLM) has successfully performed extortion without a human operator, ushering in a new era in cyberattacks that has long been expected by security experts.
Researchers at Sysdig discovered a campaign run by an “agentic threat actor” (ATA) they call JadePuffer, which exploited a flaw in an Internet-facing Langflow deployment and then pivoted to a production database server to run an adaptive and fully automated ransomware campaign, according to a recent report.
The attack unfolded across two distinct targets. First, JadePuffer exploited CVE-2025-3248, an unauthenticated remote code execution (RCE) vulnerability affecting Langflow, which is an open source tool for building AI applications.
Then JadePuffer breached its intended target: a separate Internet-exposed production server running a MySQL database and an Alibaba Nacos configuration service, according to Sysdig. Next, JadePuffer enumerated the database contents, exfiltrated selected data, deleted the database, and left an extortion note demanding payment in exchange for the stolen information.
During the attack, the machine compromised during initial access was used in the compromise of the final target, with all payloads delivered as Base64-encoded Python through the Langflow RCE endpoint.
JadePuffer Unsophisticated, Yet Adaptable
None of JadePuffer’s specific attack techniques were novel or sophisticated, says Johan Edholm, co-founder of Detectify, who calls the attack “more evolution than invention.”
“Exploiting an exposed service, harvesting credentials, moving laterally, abusing default configurations, and destroying databases are all familiar parts of the playbook,” he tells Dark Reading.
What’s remarkable, however, is that these tactics were strung together by an AI model into a complete ransomware operation against an organization’s exposed Internet-facing infrastructure, according to Sysdig. This is unlike traditional ransomware, which relies on prewritten scripts or operator intervention during key stages of an intrusion.
The most striking aspect was that “JadePuffer’s own payloads were self-narrating,” Michael Clark, Sysdig’s director of threat research, wrote in the report. “They contained natural language reasoning, target prioritization, and the kind of detailed annotations that human operators don’t often write but LLM-generated code produces reflexively.”
Sysdig observed an operation that adapted in real time, re-attempting failed steps within refined parameters in such a way that the attack went from a failed login to a working fix in 31 seconds, according to the post.
AI-Driven Ransomware Was Inevitable
Indeed, JadePuffer demonstrates how AI agents are evolving from productivity tools into autonomous offensive capabilities, representing a paradigm shift in how extortion-based attacks are carried out, according to Clark.
“An LLM agent can chain reconnaissance, credential theft, lateral movement, persistence, and destruction without the operator possessing deep expertise in any one step,” he wrote. “Tradecraft that once implied a capable human now implies a capable model.”
Since the introduction of LLMs, security experts have predicted that ransomware and other cyberattacks would inevitably be entirely generated by AI. In fact, last August, security researchers believed they discovered the first AI-driven ransomware, called PromptLock, outside of an attack scenario. The ransomware was later found to be a proof of concept developed by researchers at New York University’s Tandon School of Engineering.
JadePuffer, however, now appears to be the real deal, demonstrating how quickly AI can execute cyberattacks that can rapidly evolve in real-time beyond the automation typically seen in modern ransomware incidents.
Defensive Changes Required for AI-Driven Ransomware
Ultimately, JadePuffer demonstrates what’s possible through using AI to drive ransomware, and it serves as a warning that the volume and breadth of such campaigns will rise as agentic tooling matures, according to Sysdig. “It’s a marker of where extortion tradecraft is heading,” Clark wrote.
Edholm predicts that the earliest adopters of agentic ransomware will be actors who already know how to connect models to offensive tools, infrastructure, and stolen credentials. “As that tooling becomes packaged and reusable, it will spread to less capable operators,” he tells Dark Reading. “Criminal groups also tend to adopt new technology quickly because they are not constrained by procurement processes, compliance requirements, or other organisational bureaucracy.” Defenders also will need to move from periodic snapshots of an environment to continuous visibility, Edholm says. “A quarterly assessment can leave dangerous gaps when cloud infrastructure and Internet-facing services change every day, while an automated attacker can move from discovery to impact in minutes,” he says.
Other recommendations for organizations in the wake of the attack include patching Langflow to a release that fixes CVE-2025-3248 and ensuring code-execution/validation endpoints are not exposed to the Internet; avoiding linking provider API keys or cloud credentials to their associated AI-orchestration server environment; and hardening Nacos.
