ENCFORGE
- Type
- Ransomware
- Capabilities
- ExtortionDefense evasion
- Platforms
- Linux
ENCFORGE is a Go-based ransomware family designed to encrypt artificial intelligence and machine-learning assets. Associated with the JADEPUFFER threat actor, it has been deployed against Linux infrastructure following exploitation of CVE-2025-3248, an unauthenticated remote-code-execution vulnerability in Langflow. In the observed deployment chain, the operator abused an exposed Docker socket to create a privileged container, cross the container-host boundary, and execute the ransomware on the host.
The ransomware targets approximately 180 file extensions covering model checkpoints, model weights, vector indexes, embedding data, training datasets, and other machine-learning artifacts. Its targets include PyTorch and TensorFlow models, SafeTensors, ONNX, GGUF and GGML weights, FAISS indexes, and datasets in Parquet, Arrow, TFRecord, and NumPy formats. A command-line interface allows operators to extend the default targeting list and perform a trial scan before encryption; a companion key-generation tool is also referenced.
ENCFORGE uses region-based encryption with AES-256-CTR and wraps the per-run symmetric key with an embedded RSA-2048 public key. It renames encrypted files, drops ransom notes, terminates processes holding targeted files open, and supports resuming interrupted execution without re-encrypting completed files. The observed binary is packed with UPX and can delete itself after execution.
The operation focuses on encryption-based extortion and disruption of AI infrastructure. No data-exfiltration mechanism was identified in the recovered binary, and no data theft or associated leak site was observed in the documented campaign. Attribution to JADEPUFFER was supported by reuse of the extortion contact from the actor's earlier activity.
Associated vulnerabilities
1 CVE 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 same Langflow instance was subsequently targeted by the threat actor a second time using a compiled Go-based ransomware strain codenamed ENCFORGE.
The same Langflow instance was subsequently targeted by the threat actor a second time using a compiled Go-based ransomware strain codenamed ENCFORGE.
Techniques & procedures
28 distinct techniques documented for this family, organized by ATT&CK tactic.
Initial Access
2 techniques
Initial Access
“[It] replayed the credentials against internal databases and cache services to see what they would unlock.”
The initial JADEPUFFER operation exploited the known Langflow vulnerability CVE-2025-3248 to break in.
Execution
4 techniques
Execution
Over five minutes and 24 seconds, the operator created and revised six Python scripts through the same Langflow RCE channel until it had a working path to the host.
“In only 5 minutes and 24 seconds, the agent iterated six Python scripts through the Langflow RCE channel.”
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.
This method used privileged containers, host process access, and host filesystem mounts to escape the original container boundary.
Persistence
2 techniques
Persistence
“[It] replayed the credentials against internal databases and cache services to see what they would unlock.”
Container create requests with Privileged: true combined with PidMode: host or a bind mount of / are a high-confidence escape indicator.
Privilege Escalation
4 techniques
Privilege Escalation
the AI agent exploited the same Langflow vulnerability that was used in the earlier attack and located an exposed Docker socket that it leveraged to create a privileged escape container and deploy the ENCFORGE payload.
“[It] replayed the credentials against internal databases and cache services to see what they would unlock.”
Container create requests with Privileged: true combined with PidMode: host or a bind mount of / are a high-confidence escape indicator.
“It built a container escape mechanism to carry the binary across a namespace boundary and run it directly on the host.”
Stealth
6 techniques
Stealth
Starting with the second, the operator encoded each full script in base64 and decoded it inside an exec() call, avoiding shell-level searches for commands such as base64 -d .
“ENCFORGE is a compiled, UPX-packed Go ransomware binary.”
drops ransom notes as README , HOW_TO_DECRYPT , and README_DECRYPT , and deletes itself after running.
Following encrypting the targeted files, ENCFORGE terminated processes holding file locks and deleted itself once the encryption was complete.
“[It] replayed the credentials against internal databases and cache services to see what they would unlock.”
By launching a privileged container over the mounted Docker socket, the ransomware was moved across the container boundary through the host's proc file system and executed against the host file system outside the original container's isolation using the mounted Docker socket.
Credential Access
4 techniques
Credential Access
After gaining access and searching for cloud credentials, API tokens, and reachable internal services
“JADEPUFFER began by harvesting provider keys and cloud credentials from a single compromised host.”
After gaining access, JADEPUFFER searched for credentials
Once inside, the agent ran through the familiar reconnaissance and credential-harvest routine seen in the first campaign
Discovery
5 techniques
Discovery
“[It] probed internal services to check their reachability.”
the actor autonomously performed reconnaissance
An ENCFORGE binary was deployed in the second campaign that searched for 180 different file extensions... ENCFORGE is specifically designed to identify AI-related assets.
After gaining access and searching for cloud credentials, API tokens, and reachable internal services
The attacker generated multiple Python scripts using the compromised Langflow environment... the scripts were able to develop a method for exiting the container environment through the exposed Docker socket and executing the ransomware on the host machine.
Command and Control
1 technique
Command and Control
“[The operator] attempted to fetch the ENCFORGE binary from its command-and-control (C2) server to continue the attack.”
Impact
4 techniques
Impact
Unlike conventional ransomware targets, however, encrypted AI model artifacts cannot be restored after they are wiped.
The earlier Langflow intrusion encrypted Nacos service configuration files, leveraging MySQL's built-in AES_ENCRYPT() function. A subsequent attack used the Go-based ransomware strain ENCFORGE.
Following encrypting the targeted files, ENCFORGE terminated processes holding file locks and deleted itself once the encryption was complete.
Anti-recovery and ransom note: ENCFORGE kills processes holding file locks before encrypting data... The binary carries live Windows anti-recovery code in this Linux build: vssadmin.exe (Volume Shadow Copy deletion) and bcdedit.exe (boot recovery disable) are literal invocations in the WipeShadowCopies and DisableRecoveryActions functions.
Observables
8 indicators attributed across vendor reports, sandbox runs, and researcher write-ups. Values are masked here and available in full in Mallory.
Recent activity
17 sources tracked across advisories, community write-ups, and news. New activity surfaces here as Mallory finds it.