Skip to content

Hunting the File Writes and Process Transition After MONAI Label Exposure

A hunting sequence for suspicious datastore writes, Python startup files, training processes, and outbound traffic.

@kozhevnikova.bsky.social
Nov 21, 20265 min read

hen a vulnerable web service is found, the first question is often whether it was exploited. The MONAI Label incident provides a useful hunting sequence because the attacker had to leave traces across request handling, file creation, process launch, and outbound network activity. No single log source tells the full story.

Project-MONAI advisory GHSA-qc22-rmgc-vg94 identifies the vulnerable upload paths and affected releases. AWS reports that Cairn agents wrote a Python reverse shell to sitecustomize.py and triggered a training endpoint to launch a new Python process. The process imported the payload and connected back with root privileges on one university server.

That chain is a lead for investigation. It is not a claim that every MONAI Label system was touched.

Define the exposure window

Start with the time when the vulnerable service was reachable, the package version was installed, and authentication was disabled or weak. End the window when the service was patched, isolated, or upgraded.

Use the application's own logs to find upload requests, annotation sessions, training launches, and administrative changes. The default port is documented as 8000, but production deployments can use a proxy or another port. Do not restrict the search to one port or one hostname.

If the service was internet-facing, correlate proxy logs with application requests. If it was internal, review VPN, research-network, and container-platform logs. The search scope should reflect how the application was actually deployed.

Review the datastore request path

Look for requests to PUT /datastore/image and PUT /datastore/label. Review the image ID, label tag, extension, and path-related characters. A normal annotation workflow may use simple identifiers, but a traversal attempt can include parent-directory sequences, unusual separators, or names that resolve outside the expected directory.

Do not block the investigation on payload syntax. The server may transform or normalize the value before logging it. Preserve the raw request where available and compare it with the file path that resulted.

Search the filesystem for files created near the same time and outside the datastore. Pay special attention to Python files in import locations, application source directories, user home directories, startup configuration, SSH authorization files, and scheduler locations. Record the owner, permissions, creation time, and content hash.

Follow Python startup

The observed chain used sitecustomize.py, which Python can import during interpreter startup. That makes the file a bridge from arbitrary write to code execution. Other startup mechanisms may also be present in a Python environment.

Search for unexpected sitecustomize.py files, recent modifications to .pth files, changes to virtual-environment configuration, and Python processes with unusual import paths. Do not confuse a legitimate administrative customization with an attacker-created file. Compare the file with the base image, build record, and application owner's baseline.

If a suspicious file is found, preserve a copy before deleting it. Record its path, hash, timestamps, owner, and the process that created it. The path may show which vulnerable endpoint was used, and the content may provide network indicators.

Correlate training processes

The incident report says the agents triggered a training API endpoint to spawn a new Python process. Review the training job history around each suspicious upload. Look for new jobs, unusual command lines, unexpected working directories, or a process start that does not match a user request.

Compare the process parent, service account, environment variables, loaded modules, and outbound sockets. A training job that starts Python is normal. A training job that starts Python after a traversal-shaped upload and then contacts an unfamiliar address is a high-value correlation.

Review containers and host processes separately. A container might show a clean application root while a mounted host directory contains the written file. The process namespace and mount table are part of the evidence.

Use network indicators with care

AWS published two IP addresses present in records dated September 8 to 10, 2026: 142[.]4.98.145 and 142[.]4.107.34. Search for connections to those addresses in proxy, firewall, endpoint, and flow logs. A match is a lead, not proof by itself. Confirm the process, time, direction, and relationship to a MONAI Label request.

Do not treat absence as proof that no compromise occurred. The addresses may be blocked, the traffic may use another path, or the logs may not cover the relevant segment. The indicators are useful for corroboration and scope expansion.

Reconstruct the decision path

For each suspected host, answer four questions. Was an affected version running? Was the upload endpoint reachable by an untrusted source? Did a client submit a path that escaped the datastore? Did a subsequent process execute or import the written file?

The answers produce different outcomes. A vulnerable version with no reachable endpoint is a patch-priority issue. A reachable endpoint with no suspicious write is an exposure finding that still needs monitoring. A suspicious write without process execution is a containment event. A write followed by process execution and outbound traffic is a probable compromise that needs full incident response.

This structure is more useful than a single severity label because it tells the team what evidence is missing.

Public observations as scope input

The ZoomEye results for this event provide a public-content baseline, not a compromise list. http.body="MONAI" returned 553 observations, http.body="/datastore/image" returned 12, and the exact app and title queries returned 0 on November 21, 2026.

An organization can compare those observations with its own internet-facing inventory. It must still verify the package, endpoint, authentication, and file activity locally. The broad result cannot show which hosts run MONAI Label or which ones have the vulnerable path.

Contain and verify

If the evidence supports compromise, isolate the service at the network layer while preserving the host. Capture the filesystem, process list, memory if policy allows, application logs, proxy logs, and relevant network flows. Review the service account's recent activity and any credentials it could access.

Rotate credentials that were available to the process or stored in the application configuration. Review outbound connections and training jobs for additional activity. Remove the vulnerable service from untrusted networks and apply a fixed artifact when available.

After remediation, verify the deployed package, authentication setting, mount layout, and process account. Re-run the hunting queries against the exposure window. A clean result is meaningful only when the logs and timestamps cover the relevant period.

The MONAI Label case gives defenders a concrete chain to search for. The investigation should follow that chain while keeping observation, exposure, exploitation, and confirmed impact distinct.

References

Did you enjoy this article?

Recommend it — Standard Reader surfaces well-loved writing to more readers across the network.

Across the AtmosphereDiscussions