CVE-2026-100841: MONAI: pickle cache deserialization allows local code execution

HIGH
Published September 27, 2026
CISO Take

MONAI 1.6.0 caches medical-imaging data in a way that forces unsafe pickle deserialization (torch.load with weights_only=False). Any user who can write to a shared cache directory can plant a file that runs code as the next user who reads it. The CVSS is 7.8 (local, low privileges, no user interaction), but it is not in CISA KEV, there is no public exploit or Nuclei template, and CISA SSVC says TRACK. The EPSS score is only 0.00125, so near-term mass exploitation looks unlikely, even though the feed lists a high percentile. The risk concentrates in multi-user environments such as HPC clusters, shared GPU servers, and hospital or research workstations, where a low-privileged account could take over a data scientist's or pipeline service account. All released versions are affected and no patched version exists, so act on the workaround: give each user a private cache_dir with 0700 permissions, never use world-writable paths like /tmp, and audit existing shared caches.

Sources:NVDGitHub AdvisoryEPSSATLASvulncheck.com

What is the risk?

Medium-High in multi-tenant ML environments, Low on single-user machines. Exploitation is trivial once local write access to the cache exists (AC:L, PR:L, UI:N) and gives full confidentiality, integrity and availability impact in the victim's context. EPSS is 0.00125, there is no KEV listing, no public exploit and SSVC is TRACK, so there is no sign of active exploitation. The absence of a vendor patch extends the exposure window indefinitely. Downstream dependents and OpenSSF Scorecard data are unavailable, so blast radius is unquantified (Sin datos).

How does the attack unfold?

Initial Access
Attacker with a low-privileged local account finds a shared or world-writable MONAI cache_dir.
AML.T0012
Artifact Planting
Attacker places a malicious pickle file at a cache path that the victim's pipeline will read.
AML.T0018.002
Exploitation
The victim's PersistentDataset loads the file via torch.load(weights_only=False) or pickle.loads, executing the payload.
AML.T0011.000
Impact
Arbitrary code runs as the victim, exposing datasets, credentials and models and enabling lateral movement.
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
0.1%
chance of exploitation in 30 days
Higher than 2% of all CVEs
Exploitation Status
No known exploitation
Sophistication
Trivial

What is the attack surface?

AVACPRUISCIA
AVLocal
ACLow
PRLow
UINone
SUnchanged
CHigh
IHigh
AHigh

What should I do?

1 step
  1. 1) No patched version is available (all monai pip releases affected), so monitor the GHSA advisory GHSA-636w-j999-g7x5 for a fix and upgrade as soon as one ships. 2) Set cache_dir to a per-user directory with 0700 permissions (e.g., under the user's home); never use /tmp or other world- or group-writable paths. 3) Inventory and purge existing shared MONAI cache directories (/tmp/monai_cache, HPC scratch, ~/.cache/monai) and rebuild caches from trusted source data. 4) Where possible, avoid PersistentDataset with MetaTensors on shared storage, or use in-memory caching (CacheDataset) instead. 5) Detection: alert on writes by other UIDs into MONAI cache paths, on unexpected child processes spawned from Python training jobs, and on file ownership mismatches in cache directories. 6) Run pipelines under least-privilege service accounts and isolate users with containers or separate scratch volumes.

What does CISA's SSVC say?

DecisionTrack
Exploitationnone
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.10.3-SuppliersA.7.4-Quality of data for AI systems
NIST AI RMF
MANAGE 2.3-Responses to risks and unknown risks are managed
OWASP LLM Top 10
LLM03-Supply Chain

Frequently Asked Questions

What is CVE-2026-100841?

MONAI 1.6.0 caches medical-imaging data in a way that forces unsafe pickle deserialization (torch.load with weights_only=False). Any user who can write to a shared cache directory can plant a file that runs code as the next user who reads it. The CVSS is 7.8 (local, low privileges, no user interaction), but it is not in CISA KEV, there is no public exploit or Nuclei template, and CISA SSVC says TRACK. The EPSS score is only 0.00125, so near-term mass exploitation looks unlikely, even though the feed lists a high percentile. The risk concentrates in multi-user environments such as HPC clusters, shared GPU servers, and hospital or research workstations, where a low-privileged account could take over a data scientist's or pipeline service account. All released versions are affected and no patched version exists, so act on the workaround: give each user a private cache_dir with 0700 permissions, never use world-writable paths like /tmp, and audit existing shared caches.

Is CVE-2026-100841 actively exploited?

No confirmed active exploitation of CVE-2026-100841 has been reported, but organizations should still patch proactively.

How to fix CVE-2026-100841?

1) No patched version is available (all monai pip releases affected), so monitor the GHSA advisory GHSA-636w-j999-g7x5 for a fix and upgrade as soon as one ships. 2) Set cache_dir to a per-user directory with 0700 permissions (e.g., under the user's home); never use /tmp or other world- or group-writable paths. 3) Inventory and purge existing shared MONAI cache directories (/tmp/monai_cache, HPC scratch, ~/.cache/monai) and rebuild caches from trusted source data. 4) Where possible, avoid PersistentDataset with MetaTensors on shared storage, or use in-memory caching (CacheDataset) instead. 5) Detection: alert on writes by other UIDs into MONAI cache paths, on unexpected child processes spawned from Python training jobs, and on file ownership mismatches in cache directories. 6) Run pipelines under least-privilege service accounts and isolate users with containers or separate scratch volumes.

What systems are affected by CVE-2026-100841?

This vulnerability affects the following AI/ML architecture patterns: training pipelines, medical imaging pipelines, shared HPC/GPU clusters, data preprocessing and caching layers, multi-user notebook environments.

What is the CVSS score for CVE-2026-100841?

CVE-2026-100841 has a CVSS v3.1 base score of 7.8 (HIGH). The EPSS exploitation probability is 0.13%.

What is the AI security impact?

Affected AI Architectures

training pipelinesmedical imaging pipelinesshared HPC/GPU clustersdata preprocessing and caching layersmulti-user notebook environments

MITRE ATLAS Techniques

AML.T0010.001AI Software
AML.T0018.002Embed Malware
AML.T0050Command and Scripting Interpreter
AML.T0112.001AI Artifacts

Compliance Controls Affected

EU AI Act: Art. 15
ISO 42001: A.10.3, A.7.4
NIST AI RMF: MANAGE 2.3
OWASP LLM Top 10: LLM03

What are the technical details?

Original Advisory

In MONAI 1.6.0, PersistentDataset (monai/data/dataset.py) explicitly rejects the combination track_meta=True with weights_only=True, forcing users who cache MetaTensors (the default tensor type in MONAI >= 1.0) to run torch.load(hashfile, weights_only=False). Related cache helpers in monai/data/utils.py also call pickle.loads on cached content and derive cache keys with hashlib.md5. As a result, a local user with write access to a shared or world-writable cache_dir (e.g. /tmp/monai_cache, HPC scratch, ~/.cache/monai) can place a malicious pickle file that is deserialized the next time another user's MONAI pipeline reads the cache, resulting in arbitrary code execution in that user's context. All released versions of the monai pip package are affected; no patched version is available as of the advisory.

Exploitation Scenario

On a shared HPC or GPU server, a low-privileged user (or a compromised account) discovers a world-writable MONAI cache directory used by a research team. They craft a pickle file with a malicious __reduce__ payload and place it at a predictable cache filename, which is feasible because cache keys derive from MD5 hashes of inputs. When a data scientist or scheduled training job next loads the dataset through PersistentDataset, torch.load(weights_only=False) deserializes the file and executes the payload. The attacker obtains code execution as that user, steals SSH keys, cloud credentials and datasets, and may move laterally to a higher-privileged account or tamper with model training.

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:L/UI:N/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