Loading threat actor…
JadePuffer is an emerging, unattributed ransomware and extortion threat actor assessed as the first publicly documented agentic threat actor to direct a substantial end-to-end ransomware-style intrusion with large-language-model assistance. The actor is financially motivated and has repeatedly exploited CVE-2025-3248 in exposed Langflow deployments to gain unauthenticated remote code execution, then used the compromised AI orchestration environment as a foothold for reconnaissance, credential harvesting, internal service access, persistence, lateral movement, privilege escalation, and destructive encryption. JadePuffer has been observed targeting AI-adjacent and production infrastructure, including Langflow, MySQL, Alibaba Nacos, object storage, and internal services reachable from the compromised host. In earlier activity, the actor dumped Langflow data stores, harvested secrets such as cloud credentials, API tokens, and database credentials, reused those credentials against internal systems, and pivoted into production environments. The operation showed adaptive behavior consistent with agentic automation, including self-narrating payloads, rapid retry cycles, and on-the-fly correction of failed steps. Researchers observed hundreds of distinct payloads during one intrusion and assessed that a human operator still selected the victim, provisioned infrastructure, and likely supplied some initial credentials, while the AI-driven component materially accelerated the rest of the intrusion lifecycle. JadePuffer’s earlier campaign combined extortion with destructive actions against databases and configuration services. It encrypted and destroyed MySQL and Nacos data, created ransom artifacts, and in at least one case appeared not to preserve a recoverable decryption path, making the operation functionally destructive. Persistence was also observed through scheduled beaconing from the initial access host. The actor later evolved to deploy ENCFORGE, a purpose-built Go ransomware family designed to encrypt AI and machine-learning assets rather than only conventional enterprise files. ENCFORGE targets a broad set of AI-related artifacts, including model checkpoints, weights, vector indexes, embeddings, and training datasets. When direct payload retrieval failed, JadePuffer adapted by generating multiple Python scripts through the Langflow execution channel, abusing exposed Docker socket access, escaping from the container to the host, and executing the ransomware on the underlying system. ENCFORGE uses hybrid cryptography, kills processes holding file locks, supports resumed execution, self-deletes after execution, and drops ransom notes. Available reporting found no confirmed exfiltration capability in ENCFORGE and no associated leak site or payment portal, indicating that some observed operations were oriented more toward destructive encryption and extortion pressure than classic double extortion. JadePuffer does not currently overlap at high confidence with any established ransomware-as-a-service program, known criminal cluster, or nation-state actor. The actor is notable less for novel exploitation tradecraft than for demonstrating how agentic automation can compress a full ransomware operation against AI infrastructure into machine-speed execution. Known aliases are limited to JADEPUFFER and JadePuffer.
Mallory correlates actor tradecraft and target patterns against your stack, your sector, and your geography. See overlap before they land.
Who, where, and (when attributed) which flag flies behind the operation. Pulled from open-source reporting and Mallory's analyst review.
Sectors the actor has been observed targeting.
Attributed origin per open-source reporting.
39 distinct techniques observed across reporting, grouped by tactic. Hover any cell for the evidence excerpt; click through for MITRE's full description.
2 malware families attributed to this actor across reporting.
2 CVEs this actor has used in observed campaigns. 2 of them exploited in the wild.
This attack exploited the critical vulnerability CVE-2025-3248 that allows remote Python code execution due to a critical missing authentication vulnerability in Langflow's code-validation endpoint. Although Langflow fixed the vulnerability in version 1.3.0, the targeted server remained vulnerable even after it was publicly reported.
Exploitation de CVE-2021-29441 (bypass auth Nacos) et forge de JWT via la clé de signature par défaut de Nacos.
11 indicators attributed to this actor: domains, IPs, hashes, and other artifacts pulled from reporting. View more in app.
20 sources tracked across advisories, community write-ups, and news. New activity surfaces here as Mallory finds it.
Conducting a ransomware campaign against AI environments by exploiting a Langflow vulnerability and deploying ENCFORGE to encrypt AI model weights, vector indexes, and training data, with apparent emphasis on destruction/disruption rather than double extortion.
Targeted exposed AI and product-lifecycle platforms for encryption, data theft, and extortion.
A ransomware operation referenced as using an AI agent to automate the full intrusion lifecycle, including reconnaissance, credential theft, lateral movement, privilege escalation, and data encryption.
Agentic ransomware operations targeting AI assets, including trained models, datasets, and related artifacts, by exploiting Langflow and abusing exposed Docker sockets to deploy ENCFORGE for encryption.
Match sector + geo + tech-stack targeting against your real footprint.
Every observed MITRE ATT&CK technique, grouped by tactic.
Families this actor is known to deploy, with IOCs and behavior.
CVEs this actor has used in known campaigns.
YARA, Sigma, Snort, and vendor rules, auto-deployed to your SIEM.
Domains, IPs, and hashes tied to this actor, refreshed continuously.