CVE-2026-0897: Keras Resource Exhaustion enables DoS

CISO Take

If your ML pipelines or model serving infrastructure load .keras files from external, user-controlled, or shared repository sources, patch Keras to 3.12.1 now. A crafted weight archive can crash the Python interpreter with zero authentication required, taking down inference workers or training jobs. If immediate patching is blocked, enforce strict allowlisting of model sources and apply container memory limits to bound blast radius.

What is the risk?

MEDIUM-HIGH for organizations running public model evaluation endpoints, automated fine-tuning pipelines accepting external checkpoints, or transfer learning workflows pulling from shared repositories. LOW for air-gapped environments loading only internally-signed, origin-verified weights. EPSS of 0.00029 indicates minimal active exploitation today, but the attack primitive is trivially reproducible by any attacker who can deliver a crafted file to a pipeline — no ML expertise required beyond understanding HDF5 metadata structure.

What systems are affected?

Package	Ecosystem	Vulnerable Range	Patched
Keras	pip	>= 3.0.0, <= 3.12.0	`3.12.1`
64.1K OpenSSF 7.1 1.6K dependents Pushed 5d ago 48% patched ~32d to patch Full package profile →

Do you use Keras? You're affected.

How severe is it?

CVSS 3.1

N/A

EPSS

0.4%

chance of exploitation in 30 days

Higher than 28% of all CVEs

Source: EPSS v3 — FIRST.org

Exploitation Status

No known exploitation

Sophistication

Trivial

What should I do?

1 step

1) Patch immediately: pip install 'keras>=3.12.1'. Verify with pip show keras. 2) If patching is blocked, restrict model loading to cryptographically verified, internally-hosted artifacts only — reject any externally-sourced .keras or .h5 files at the pipeline boundary. 3) Apply container/cgroup memory limits on all ML serving and training pods to prevent host-level memory exhaustion from a single crashing process. 4) Add file size validation and HDF5 shape metadata inspection before invoking keras.saving.load_model() or equivalent. 5) Detection: alert on abnormal RSS/VSZ spikes in Python ML processes, or sudden OOMKilled pod events correlated with model load operations. 6) Audit all publicly-accessible endpoints that trigger weight loading — model fine-tuning APIs, evaluation services, and any user-facing upload flows.

What does CISA's SSVC say?

Decision Track

Exploitation none

Automatable No

Technical Impact partial

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

How is it classified?

DoS Supply Chain Framework Model AML.T0010.001 - AI Software AML.T0011.000 - Unsafe AI Artifacts AML.T0029 - Denial of AI Service AML.T0049 - Exploit Public-Facing Application

Which compliance frameworks are affected?

This CVE is relevant to:

EU AI Act

Article 15 - Accuracy, robustness and cybersecurity

ISO 42001

A.6.1.5 - AI system security A.6.2.6 - AI system security

NIST AI RMF

MANAGE 2.4 - Residual risks to AI are managed and monitored MS-2.5 - Manage AI risks on an ongoing basis

OWASP LLM Top 10

LLM04 - Model Denial of Service LLM04:2025 - Model Denial of Service

Frequently Asked Questions

What is CVE-2026-0897?

If your ML pipelines or model serving infrastructure load .keras files from external, user-controlled, or shared repository sources, patch Keras to 3.12.1 now. A crafted weight archive can crash the Python interpreter with zero authentication required, taking down inference workers or training jobs. If immediate patching is blocked, enforce strict allowlisting of model sources and apply container memory limits to bound blast radius.

Is CVE-2026-0897 actively exploited?

No confirmed active exploitation of CVE-2026-0897 has been reported, but organizations should still patch proactively.

How to fix CVE-2026-0897?

1) Patch immediately: pip install 'keras>=3.12.1'. Verify with pip show keras. 2) If patching is blocked, restrict model loading to cryptographically verified, internally-hosted artifacts only — reject any externally-sourced .keras or .h5 files at the pipeline boundary. 3) Apply container/cgroup memory limits on all ML serving and training pods to prevent host-level memory exhaustion from a single crashing process. 4) Add file size validation and HDF5 shape metadata inspection before invoking keras.saving.load_model() or equivalent. 5) Detection: alert on abnormal RSS/VSZ spikes in Python ML processes, or sudden OOMKilled pod events correlated with model load operations. 6) Audit all publicly-accessible endpoints that trigger weight loading — model fine-tuning APIs, evaluation services, and any user-facing upload flows.

What systems are affected by CVE-2026-0897?

This vulnerability affects the following AI/ML architecture patterns: training pipelines, model serving, transfer learning workflows, MLOps/CI-CD pipelines, model evaluation platforms.

What is the CVSS score for CVE-2026-0897?

No CVSS score has been assigned yet.

What is the AI security impact?

Affected AI Architectures

training pipelinesmodel servingtransfer learning workflowsMLOps/CI-CD pipelinesmodel evaluation platforms

MITRE ATLAS Techniques

AML.T0010.001 AI Software

AML.T0011.000 Unsafe AI Artifacts

AML.T0029 Denial of AI Service

AML.T0049 Exploit Public-Facing Application

Compliance Controls Affected

EU AI Act: Article 15

ISO 42001: A.6.1.5, A.6.2.6

NIST AI RMF: MANAGE 2.4, MS-2.5

OWASP LLM Top 10: LLM04, LLM04:2025

What are the technical details?

Original Advisory

Allocation of Resources Without Limits or Throttling in the HDF5 weight loading component in Google Keras 3.0.0 through 3.12.0 and 3.13.0 on all platforms allows a remote attacker to cause a Denial of Service (DoS) through memory exhaustion and a crash of the Python interpreter via a crafted .keras archive containing a valid model.weights.h5 file whose dataset declares an extremely large shape.

Exploitation Scenario

An adversary targeting an organization running an automated model evaluation platform or transfer learning pipeline crafts a valid .keras archive. The archive contains a well-formed model.weights.h5 with a legitimate HDF5 header but declares a dataset shape with extreme dimensions (e.g., [2147483647, 2147483647]). When the pipeline invokes keras.saving.load_model() or loads the weights file, the HDF5 component attempts to pre-allocate memory proportional to the declared shape before reading actual data. System RAM is exhausted within seconds, the Python interpreter crashes via an unhandled MemoryError, and the inference worker or training job is terminated. In containerized ML workloads without memory limits, this can cascade to OOMKill events affecting co-located services. No authentication, no ML expertise, and no network-level access beyond file delivery to the target pipeline are required.

Weaknesses (CWE)

CWE-770 Allocation of Resources Without Limits or Throttling Primary

CWE-770 — Allocation of Resources Without Limits or Throttling: The product allocates a reusable resource or group of resources on behalf of an actor without imposing any intended restrictions on the size or number of resources that can be allocated.

[Requirements] Clearly specify the minimum and maximum expectations for capabilities, and dictate which behaviors are acceptable when resource allocation reaches limits.
[Architecture and Design] Limit the amount of resources that are accessible to unprivileged users. Set per-user limits for resources. Allow the system administrator to define these limits. Be careful to avoid CWE-410.

Source: MITRE CWE corpus.