CVE-2026-92786: LightGBM OOB write — HIGH

Q: Is CVE-2026-92786 actively exploited?

Proof-of-concept exploit code is publicly available for CVE-2026-92786, increasing the risk of exploitation.

Q: What systems are affected by CVE-2026-92786?

This vulnerability affects the following AI/ML architecture patterns: model explainability pipelines, training pipelines, model serving, notebook environments, MLOps model validation.

Q: What is the CVSS score for CVE-2026-92786?

CVE-2026-92786 has a CVSS v3.1 base score of 7.8 (HIGH). The EPSS exploitation probability is 0.19%.

CISO Take

LightGBM through 4.7.0 does not validate child and split array values when parsing text model files, so a crafted model can trigger out-of-bounds memory writes at attacker-chosen offsets during SHAP feature-contribution computation (CWE-787). The CVSS is 7.8 (local, low complexity, no privileges, user interaction required), and the realistic trigger is a team loading and explaining an untrusted model, for example one downloaded from a public hub or supplied by a vendor or contractor. Exploitation likelihood is modest: EPSS is 0.00186 (92nd percentile), the CVE is not in CISA KEV, there is no public exploit or Nuclei template, and the CISA SSVC decision is TRACK. Dependent counts and package risk data are unavailable, but LightGBM is a very widely used gradient boosting library, so inventory where it is deployed. Treat model files as untrusted code-equivalent input: upgrade once a fixed release past 4.7.0 is available, and until then do not run SHAP or pred_contrib on models from unverified sources.

Sources:NVDEPSSATLASVulnCheckGitHub

What is the risk?

Medium-high severity but low near-term exploitation risk. Impact is high for confidentiality, integrity and availability (memory corruption with attacker-controlled write offsets, which could lead to code execution). Exploitation needs local access and a victim loading a malicious model and calling SHAP/contribution prediction. There is no known public exploit, no KEV listing, and EPSS is low in absolute terms, although the percentile is high. Exposure is greatest in data science workstations, notebooks, CI pipelines and model-explainability services that ingest third-party models.

How does the attack unfold?

Model Delivery

Attacker crafts a LightGBM text model with invalid child/split node references and distributes it via a public repository or supplier channel.

AML.T0058

User Execution

Victim or an automated pipeline loads the malicious model file and invokes SHAP/feature contribution prediction.

AML.T0011.000

Exploitation

Unvalidated node indices cause out-of-bounds writes at attacker-chosen offsets in the leaf_depth_ buffer, corrupting memory.

Impact

Memory corruption leads to a crash or potential arbitrary code execution, exposing data, credentials and other models in the environment.

AML.T0112.001

Model Delivery

Attacker crafts a LightGBM text model with invalid child/split node references and distributes it via a public repository or supplier channel.

AML.T0058

User Execution

Victim or an automated pipeline loads the malicious model file and invokes SHAP/feature contribution prediction.

AML.T0011.000

Exploitation

Unvalidated node indices cause out-of-bounds writes at attacker-chosen offsets in the leaf_depth_ buffer, corrupting memory.

Impact

Memory corruption leads to a crash or potential arbitrary code execution, exposing data, credentials and other models in the environment.

AML.T0112.001

What systems are affected?

Package	Ecosystem	Vulnerable Range	Patched
LightGBM	pip	—	No patch
18.9KOpenSSF 6.21.1K dependentsPushed 5d ago33% patched~721d to patchFull package profile →

Do you use LightGBM? You're affected.

How severe is it?

CVSS 3.1

7.8 / 10

EPSS

0.2%

chance of exploitation in 30 days

Higher than 8% of all CVEs

Source: EPSS v3 — FIRST.org

Exploitation Status

Exploit Available

Exploitation: MEDIUM

Sophistication

Advanced

Exploitation Confidence

medium

○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?

AVLocal

ACLow

PRNone

UIRequired

SUnchanged

CHigh

IHigh

AHigh

What should I do?

1 step

1) Inventory LightGBM usage (pip lightgbm, plus embedded builds in other tools) and identify versions <= 4.7.0. 2) Upgrade to a fixed release once published by the lightgbm-org maintainers; track GitHub issue #7357 and the VulnCheck advisory. 3) Until patched, only load models from trusted, integrity-verified sources (signed artifacts, hash pinning, internal registry). 4) Avoid SHAP/pred_contrib on untrusted models, or run it in a sandboxed, low-privilege container with no credentials or network access. 5) Detect: monitor for crashes or segfaults in explainability jobs, and alert on model files sourced from outside approved registries. 6) Add model-file provenance checks to MLOps pipelines.

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?

Code ExecutionSupply ChainDoSFrameworkModelAML.T0010.003- Model AML.T0011.000- Unsafe AI Artifacts AML.T0018.002- Embed Malware AML.T0058- Publish Poisoned Models

Which compliance frameworks are affected?

This CVE is relevant to:

EU AI Act

Article 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 previously unknown risks and supply chain risks

OWASP LLM Top 10

LLM03-Supply Chain

Frequently Asked Questions

What is CVE-2026-92786?

LightGBM through 4.7.0 does not validate child and split array values when parsing text model files, so a crafted model can trigger out-of-bounds memory writes at attacker-chosen offsets during SHAP feature-contribution computation (CWE-787). The CVSS is 7.8 (local, low complexity, no privileges, user interaction required), and the realistic trigger is a team loading and explaining an untrusted model, for example one downloaded from a public hub or supplied by a vendor or contractor. Exploitation likelihood is modest: EPSS is 0.00186 (92nd percentile), the CVE is not in CISA KEV, there is no public exploit or Nuclei template, and the CISA SSVC decision is TRACK. Dependent counts and package risk data are unavailable, but LightGBM is a very widely used gradient boosting library, so inventory where it is deployed. Treat model files as untrusted code-equivalent input: upgrade once a fixed release past 4.7.0 is available, and until then do not run SHAP or pred_contrib on models from unverified sources.

Is CVE-2026-92786 actively exploited?

Proof-of-concept exploit code is publicly available for CVE-2026-92786, increasing the risk of exploitation.

How to fix CVE-2026-92786?

1) Inventory LightGBM usage (pip lightgbm, plus embedded builds in other tools) and identify versions <= 4.7.0. 2) Upgrade to a fixed release once published by the lightgbm-org maintainers; track GitHub issue #7357 and the VulnCheck advisory. 3) Until patched, only load models from trusted, integrity-verified sources (signed artifacts, hash pinning, internal registry). 4) Avoid SHAP/pred_contrib on untrusted models, or run it in a sandboxed, low-privilege container with no credentials or network access. 5) Detect: monitor for crashes or segfaults in explainability jobs, and alert on model files sourced from outside approved registries. 6) Add model-file provenance checks to MLOps pipelines.

What systems are affected by CVE-2026-92786?

This vulnerability affects the following AI/ML architecture patterns: model explainability pipelines, training pipelines, model serving, notebook environments, MLOps model validation.

What is the CVSS score for CVE-2026-92786?

CVE-2026-92786 has a CVSS v3.1 base score of 7.8 (HIGH). The EPSS exploitation probability is 0.19%.

What is the AI security impact?

Affected AI Architectures

model explainability pipelinestraining pipelinesmodel servingnotebook environmentsMLOps model validation

MITRE ATLAS Techniques

AML.T0010.003Model

AML.T0011.000Unsafe AI Artifacts

AML.T0018.002Embed Malware

AML.T0058Publish Poisoned Models

Compliance Controls Affected

EU AI Act: Article 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

LightGBM through 4.7.0 fails to validate child and split array values when parsing text models, allowing attackers to write out-of-bounds memory during SHAP prediction. Attackers can craft malicious model files with invalid node references that trigger out-of-bounds writes at attacker-chosen offsets in the leaf_depth_ buffer during feature contribution computation.

Exploitation Scenario

An attacker publishes a plausible-looking LightGBM text model on a public repository, or sends it to a target as part of a vendor deliverable or shared benchmark. The model contains invalid child/split node references. A data scientist or automated model-validation pipeline loads the model and runs SHAP to explain predictions. The malformed node indices cause out-of-bounds writes at attacker-chosen offsets in leaf_depth_, corrupting memory and potentially hijacking control flow to run code with the privileges of the analyst or pipeline, from which the attacker can steal data, credentials or tamper with other models.

Weaknesses (CWE)

CWE-787Out-of-bounds WritePrimary CWE-787Out-of-bounds Write

CWE-787 — Out-of-bounds Write: The product writes data past the end, or before the beginning, of the intended buffer.

[Requirements] Use a language that does not allow this weakness to occur or provides constructs that make this weakness easier to avoid. For example, many languages that perform their own memory management, such as Java and Perl, are not subject to buffer overflows. Other languages, such as Ada and C#, typically provide overflow protection, but the protection can be disabled by the programmer. Be wary that a language's interface to native code may still be subject to overflows, even if the language itself is theoretically safe.
[Architecture and Design] Use a vetted library or framework that does not allow this weakness to occur or provides constructs that make this weakness easier to avoid. Examples include the Safe C String Library (SafeStr) by Messier and Viega [REF-57], and the Strsafe.h library from Microsoft [REF-56]. These libraries provide safer versions of overflow-prone string-handling functions.

Source: MITRE CWE corpus.