JADEPUFFER
- Type
- Ransomware
- Capabilities
- Initial accessReconnaissanceScanningCredential theftExfiltrationPersistenceLateral movementExtortion
- Delivery
- Landing page
- Platforms
- Linux
JADEPUFFER is an AI-agent-assisted ransomware operation associated with the financially motivated threat actor Storm-3168. Publicly documented in July 2026, it uses large language model assistance to coordinate exploitation, reconnaissance, credential harvesting, persistence, lateral movement, and data encryption. Initial claims that the operation was fully autonomous were subsequently qualified: human operators remained involved in target selection, infrastructure preparation, and attack oversight.
The operation exploits CVE-2025-3248, an unauthenticated remote code execution vulnerability in Langflow, to execute Python payloads on exposed AI workflow servers. Following compromise, it enumerates systems and internal services and harvests secrets, including cloud credentials, LLM-provider API keys, and database connection information. Observed activity included accessing MinIO with default credentials, establishing scheduled-task persistence, and pivoting to a production environment running MySQL and Alibaba Nacos.
JADEPUFFER abused a default Nacos signing key to forge authentication tokens and used database access to create a backdoor administrator account. It encrypted more than 1,300 configuration records, removed original database tables, and left a ransom demand. Its agent-assisted workflow adapted payloads after execution and authentication failures. Later activity included deployment of ENCFORGE, a locker targeting AI and machine-learning assets such as model checkpoints, vector databases, embedding indexes, and training data. Its targets include exposed enterprise AI infrastructure and associated production data services.
Associated vulnerabilities
3 CVEs Mallory has correlated with this family across public research and vendor advisories. Each card links to the full Mallory page for that vulnerability.
Associated threat actors
2 distinct threat actors Mallory has associated with this family across public reporting. Open in Mallory to see the full evidence chain and overlapping campaigns.
The activity is linked to JADEPUFFER agentic ransomware operations, documented earlier in a different intrusion.
Sysdig researchers were able to detect the campaign by analyzing an attack linked to the JadePuffer threat actor that exploited a critical vulnerability in Langflow to gain initial access.
Techniques & procedures
51 distinct techniques documented for this family, organized by ATT&CK tactic.
Reconnaissance
1 technique
Reconnaissance
The agent found a year-old vulnerability in an internet-facing Langflow instance and enumerated the environment.
Resource Development
1 technique
Resource Development
Для MITRE ATT&CK это маппируется на технику T1588.007 (Artificial Intelligence) из тактики Resource Development
Initial Access
5 techniques
Initial Access
Агент обнаружил MinIO на 127.0.0.1:9000 и провёл полный цикл перечисления с дефолтными credentials
probing various services using default credentials. During a MinIO object storage enumeration, the agent used default credentials (minioadmin:minioadmin)
“Storm-3168 used compromised cloud identities” and “a stolen application credential can give an intruder broad control”; two compromised Azure service principals were used in the tenant.
the second, moving laterally to the ultimate target—an independent production database server.
CVE-2025-3248 - отсутствие аутентификации на эндпоинте /api/v1/validate/code , позволяющее удалённому неаутентифицированному атакующему выполнить произвольный Python через crafted HTTP-запрос.
Execution
3 techniques
Execution
Persistence - через crontab с beacon каждые 30 минут на C2-сервер 45.131.66[.]106:4444 .
There were no new MITRE attack techniques. Seven of eight operations ran T1059, Command & Scripting Interpreter, the single most ordinary technique in the framework.
JADEPUFFER использовал именно эту уязвимость, доставляя все payload'ы как Base64-encoded Python через RCE-эндпоинт.
Persistence
7 techniques
Persistence
Persistence - через crontab с beacon каждые 30 минут на C2-сервер 45.131.66[.]106:4444 .
Агент обнаружил MinIO на 127.0.0.1:9000 и провёл полный цикл перечисления с дефолтными credentials
probing various services using default credentials. During a MinIO object storage enumeration, the agent used default credentials (minioadmin:minioadmin)
“Storm-3168 used compromised cloud identities” and “a stolen application credential can give an intruder broad control”; two compromised Azure service principals were used in the tenant.
the second, moving laterally to the ultimate target—an independent production database server.
Инъекция backdoor-администратора напрямую в backing MySQL через root-доступ к БД.
forging JWTs using the public default signing key, and directly injecting a backdoor admin account
Privilege Escalation
6 techniques
Privilege Escalation
Persistence - через crontab с beacon каждые 30 минут на C2-сервер 45.131.66[.]106:4444 .
Privilege Escalation Exploited Nacos (CVE-2021-29441) to bypass authentication, forge JWTs, and create administrative access.
Агент обнаружил MinIO на 127.0.0.1:9000 и провёл полный цикл перечисления с дефолтными credentials
probing various services using default credentials. During a MinIO object storage enumeration, the agent used default credentials (minioadmin:minioadmin)
“Storm-3168 used compromised cloud identities” and “a stolen application credential can give an intruder broad control”; two compromised Azure service principals were used in the tenant.
Параллельно с этим ИИ искал возможность осуществить побег из контейнера
Stealth
5 techniques
Stealth
...before staging the output to local files before reviewing and deleting them.
Агент обнаружил MinIO на 127.0.0.1:9000 и провёл полный цикл перечисления с дефолтными credentials
probing various services using default credentials. During a MinIO object storage enumeration, the agent used default credentials (minioadmin:minioadmin)
“Storm-3168 used compromised cloud identities” and “a stolen application credential can give an intruder broad control”; two compromised Azure service principals were used in the tenant.
Tous les payloads étaient délivrés en Python encodé en Base64 via cet endpoint RCE.
Defense Impairment
2 techniques
Defense Impairment
forging JWTs using the public default signing key, and directly injecting a backdoor admin account
# T1562.001 — Impair Defenses: Disable or Modify Tools (Defense Evasion)
Credential Access
10 techniques
Credential Access
Dubbed JadePuffer, the agent was used for harvesting and reusing credentials...
During the ransomware attack, JADEPUFFER hit an initial login attempt failure. The agent diagnosed the cause and issued a corrected payload within 31 seconds
The marimo case is the clean demonstration: an ATA gained entry through an ordinary CVE, then composed the entire post-exploitation chain live — credential harvesting, an AWS Secrets Manager call...
The same service principal made an inventory request for Azure Storage Accounts and sent more than 30 successful ListKeys requests, asking ARM to return each storage account's access keys.
Collecte de secrets ... credentials cloud ... credentials base de données ... extraction de credentials.json et fichiers .env
Collecte de secrets : clés API LLM (OpenAI, Anthropic, DeepSeek, Gemini)
Параллельно - sweep credentials по категориям: API-ключи LLM-провайдеров, облачные credentials (AWS, GCP, Azure, Alibaba, Aliyun, Tencent, Huawei), криптовалютные кошельки, конфигурации баз данных.
forging JWTs using the public default signing key, and directly injecting a backdoor admin account
подделывал JWT с помощью стандартного ключа подписи
forged a JWT with the default signing key
Discovery
9 techniques
Discovery
Activities included: Enumerating hosts Discovering services Identifying accessible resources Mapping network relationships
Сканирование внутреннего адресного пространства с пробингом databases, object storage, secret stores.
enumerating host information (system identification, kernel version, hostname, network interfaces, and running processes)
AI-агент провёл System Information Discovery (T1082): id , uname -a , hostname , перечисление сетевых интерфейсов и процессов.
Reconnaissance système (id, uname, hostname, interfaces réseau, processus) ... extraction de credentials.json et fichiers .env
Collecte de secrets ... credentials cloud ... credentials base de données
“the first compromised identity spent about 15 hours and 30 minutes performing more than 300 successful read operations. It listed virtual machines, subscriptions, resource groups, and other resources.”
“the second identity checked virtual machines and groups across two subscriptions” and “examined App Service configuration stores.”
During a MinIO object storage enumeration, the agent used default credentials (minioadmin:minioadmin) to list all buckets
Lateral Movement
2 techniques
Lateral Movement
Moved laterally to a production server running MySQL and Alibaba Nacos.
it chained together known vulnerabilities, exposed services and default credentials before adapting its behaviour in real time when obstacles emerged
Collection
3 techniques
Collection
целевой GetObject на credentials.json и .env
...before staging the output to local files before reviewing and deleting them.
Dumped the Langflow PostgreSQL database.
Command and Control
3 techniques
Command and Control
http://45.131.66.106:4444/beacon
beaconant toutes les 30 minutes vers 45.131.66.106:4444 ... URLs : http://45.131.66.106:4444/beacon
The exploitation of CVE-2026-33017 is followed by the deployment of payloads designed to fetch a second-stage downloader responsible for delivering additional malware.
Exfiltration
2 techniques
Exfiltration
Installation d’une persistance crontab beaconant toutes les 30 minutes vers 45.131.66.106:4444
A human still set up and pointed the operation and provisioned the infrastructure behind it, the command-and-control server, the staging server used for the stolen data...
Impact
6 techniques
Impact
“Over roughly seven minutes, Storm-3168 made more than 100 attempts to delete storage accounts, with most targeted accounts successfully removed. It also deleted a Key Vault, Function App, and App Service plan.”
JadePuffer, qui serait une attaque par rançongiciel entièrement automatisée, chaque étape étant gérée par l’IA du début à la fin.
“It separately targeted Site Recovery and Azure Backup protection locks” and targeted “storage accounts with terraform and backup-themed names.”
création d’une table de rançon README_RANSOM
before leaving a ransom note
...encrypted a production database before demanding a bitcoin ransom.
Observables
5 indicators attributed across vendor reports, sandbox runs, and researcher write-ups. Values are masked here and available in full in Mallory.
Recent activity
51 sources tracked across advisories, community write-ups, and news. New activity surfaces here as Mallory finds it.