CVE-2026-100845: MONAI: unsafe pickle deserialization in NumpyReader RCE

HIGHPoC AVAILABLECISA: ATTEND
Published September 27, 2026
CISO Take

MONAI versions before 1.6.0 load .npy and .npz files through NumpyReader with numpy.load(allow_pickle=True) hardcoded, so any crafted file carrying a pickle payload executes arbitrary code as soon as it passes through the standard data pipeline. The CVSS of 7.8 is driven by local vector and required user interaction, but in practice the trigger is simply ingesting a dataset or shared file, which is routine in medical imaging research and training workflows. Exploitation likelihood is currently low in absolute terms (EPSS 0.00144), yet that score sits in the top 97th percentile, there is no public exploit, no Nuclei template and no CISA KEV listing, and CISA SSVC rates it ATTEND; the technique itself (malicious pickle) is trivial and well understood. Upgrade MONAI to 1.6.0 or later, and until then block untrusted .npy/.npz inputs and scan them with picklescan or fickling before they reach training or inference jobs.

Sources:NVDEPSSGitHub AdvisoryATLASvulncheck.com

What is the risk?

Medium-High. Impact is full confidentiality, integrity and availability loss on the host running the pipeline (C:H/I:H/A:H), with code execution under the identity of the training or inference process. Exploitability is easy once a victim loads an attacker-controlled file (AC:L, PR:N), but it requires user interaction and local delivery of the file. No known active exploitation, no public PoC, not in KEV. Exposure depends on whether pipelines ingest externally sourced data (shared datasets, collaborator uploads, public repositories). Healthcare and research environments often run these jobs on GPU hosts with access to PHI, credentials and cloud storage, which raises business impact. Dependent count and scorecard data are unknown, so blast radius could not be quantified.

How does the attack unfold?

Payload Crafting
Attacker builds a .npy/.npz file with an embedded pickle payload that runs code on deserialization.
AML.T0017
Delivery
Malicious file is distributed via a shared dataset, repository or collaborator exchange to the victim's data pipeline.
AML.T0010.002
Execution
Victim loads the file through MONAI NumpyReader, where numpy.load with allow_pickle=True executes the payload.
AML.T0011.000
Impact
Attacker gains code execution on the ML host, stealing data, credentials and model weights, and pivoting further.
AML.T0112.001

What systems are affected?

PackageEcosystemVulnerable RangePatched
MONAIpip—No patch
8.8KOpenSSF 7.8129 dependentsPushed 4d ago41% patched~22d to patchFull package profile →

Do you use MONAI? You're affected.

How severe is it?

CVSS 3.1
7.8 / 10
EPSS
0.2%
chance of exploitation in 30 days
Higher than 10% of all CVEs
Exploitation Status
Exploit Available
Exploitation: MEDIUM
Sophistication
Trivial
Exploitation Confidence
medium
○CISA SSVC: Public PoC
○Public PoC indexed (trickest/cve)
Composite signal derived from CISA KEV, VulnCheck KEV, CISA SSVC, EPSS, Metasploit, Exploit-DB, trickest/cve, Nuclei templates, and inthewild.io exploitation reports.

What is the attack surface?

AVACPRUISCIA
AVLocal
ACLow
PRNone
UIRequired
SUnchanged
CHigh
IHigh
AHigh

What should I do?

1 step
  1. 1) Upgrade MONAI to 1.6.0 or later and verify with pip show monai in all environments, containers and notebooks. 2) Inventory pipelines that use NumpyReader or LoadImage on .npy/.npz inputs. 3) Until patched, accept only trusted data sources, convert inputs to safer formats (NIfTI, DICOM, or numpy arrays saved without object dtype), and reject files that require pickle. 4) Scan incoming numpy files with picklescan or fickling before ingestion. 5) Run data loading jobs in sandboxed, least-privilege containers with no access to production credentials and restricted egress. 6) Detection: monitor for unexpected child processes, network connections or file writes from training/inference Python processes, and alert on numpy.load calls with allow_pickle=True in code review/SAST. 7) Track the GHSA-wg9g-w2j2-8pgr advisory for updates.

What does CISA's SSVC say?

DecisionAttend
Exploitationpoc
AutomatableNo
Technical Impacttotal

Source: CISA Vulnrichment (SSVC v2.0). Decision based on the CISA Coordinator decision tree.

How is it classified?

Which compliance frameworks are affected?

This CVE is relevant to:

EU AI Act
Art. 15-Accuracy, robustness and cybersecurity
ISO 42001
A.7.4-Quality of data for AI systems
NIST AI RMF
MANAGE 2.3-Responding to unknown risks and managing AI supply chain risks
OWASP LLM Top 10
LLM03-Supply Chain

Frequently Asked Questions

What is CVE-2026-100845?

MONAI versions before 1.6.0 load .npy and .npz files through NumpyReader with numpy.load(allow_pickle=True) hardcoded, so any crafted file carrying a pickle payload executes arbitrary code as soon as it passes through the standard data pipeline. The CVSS of 7.8 is driven by local vector and required user interaction, but in practice the trigger is simply ingesting a dataset or shared file, which is routine in medical imaging research and training workflows. Exploitation likelihood is currently low in absolute terms (EPSS 0.00144), yet that score sits in the top 97th percentile, there is no public exploit, no Nuclei template and no CISA KEV listing, and CISA SSVC rates it ATTEND; the technique itself (malicious pickle) is trivial and well understood. Upgrade MONAI to 1.6.0 or later, and until then block untrusted .npy/.npz inputs and scan them with picklescan or fickling before they reach training or inference jobs.

Is CVE-2026-100845 actively exploited?

Proof-of-concept exploit code is publicly available for CVE-2026-100845, increasing the risk of exploitation.

How to fix CVE-2026-100845?

1) Upgrade MONAI to 1.6.0 or later and verify with pip show monai in all environments, containers and notebooks. 2) Inventory pipelines that use NumpyReader or LoadImage on .npy/.npz inputs. 3) Until patched, accept only trusted data sources, convert inputs to safer formats (NIfTI, DICOM, or numpy arrays saved without object dtype), and reject files that require pickle. 4) Scan incoming numpy files with picklescan or fickling before ingestion. 5) Run data loading jobs in sandboxed, least-privilege containers with no access to production credentials and restricted egress. 6) Detection: monitor for unexpected child processes, network connections or file writes from training/inference Python processes, and alert on numpy.load calls with allow_pickle=True in code review/SAST. 7) Track the GHSA-wg9g-w2j2-8pgr advisory for updates.

What systems are affected by CVE-2026-100845?

This vulnerability affects the following AI/ML architecture patterns: training pipelines, data preprocessing pipelines, model serving, medical imaging ML workflows, research notebooks.

What is the CVSS score for CVE-2026-100845?

CVE-2026-100845 has a CVSS v3.1 base score of 7.8 (HIGH). The EPSS exploitation probability is 0.20%.

What is the AI security impact?

Affected AI Architectures

training pipelinesdata preprocessing pipelinesmodel servingmedical imaging ML workflowsresearch notebooks

MITRE ATLAS Techniques

AML.T0010.002Data
AML.T0011.000Unsafe AI Artifacts
AML.T0018.002Embed Malware
AML.T0112.001AI Artifacts

Compliance Controls Affected

EU AI Act: Art. 15
ISO 42001: A.7.4
NIST AI RMF: MANAGE 2.3
OWASP LLM Top 10: LLM03

What are the technical details?

Original Advisory

MONAI before 1.6.0 contains an unsafe deserialization vulnerability in the NumpyReader class that unconditionally uses numpy.load with allow_pickle=True when loading .npy and .npz files. Attackers can craft malicious .npy files with pickle payloads that execute arbitrary code when loaded through MONAI's standard data pipeline.

Exploitation Scenario

An attacker publishes or shares a seemingly legitimate medical imaging dataset (or sends a collaborator an .npy preprocessing artifact) in which one or more .npy files embed a malicious pickle payload. A data scientist at the target organization points a MONAI training or preprocessing pipeline at the dataset. NumpyReader calls numpy.load with allow_pickle=True, the payload runs during deserialization and spawns a reverse shell on the GPU workstation. The attacker then harvests cloud credentials, steals patient data and model weights, and pivots into the broader ML infrastructure.

Weaknesses (CWE)

CWE-502 — Deserialization of Untrusted Data: The product deserializes untrusted data without sufficiently ensuring that the resulting data will be valid.

  • [Architecture and Design, Implementation] If available, use the signing/sealing features of the programming language to assure that deserialized data has not been tainted. For example, a hash-based message authentication code (HMAC) could be used to ensure that data has not been modified.
  • [Implementation] When deserializing data, populate a new object rather than just deserializing. The result is that the data flows through safe input validation and that the functions are safe.

Source: MITRE CWE corpus.

CVSS Vector

CVSS:3.1/AV:L/AC:L/PR:N/UI:R/S:U/C:H/I:H/A:H

Timeline

Published
September 27, 2026
Last Modified
September 30, 2026
First Seen
September 27, 2026

Related Vulnerabilities