CVE-2022-23583 — MEDIUM (CVSS 6.5) AI Security Vulnerability

CISO Take

A low-privileged attacker can crash TensorFlow inference services by supplying a malicious SavedModel with mismatched dtype fields, causing CHECK assertion failures. Any platform that loads externally-supplied or user-provided SavedModels is directly exposed. Patch to TF 2.8.0, 2.7.1, 2.6.3, or 2.5.3 immediately and restrict who can push models to your serving infrastructure.

Risk Assessment

Medium operational risk. CVSS 6.5 understates the practical impact for multi-tenant ML platforms: low complexity, low privileges, and network delivery make this trivially weaponizable for service disruption. Confidentiality and integrity are unaffected, but availability impact is high for shared inference endpoints. Organizations running internal-only, air-gapped model serving with controlled model provenance face significantly lower exposure.

Affected Systems

Package	Ecosystem	Vulnerable Range	Patched
tensorflow	pip	—	No patch
195.0K OpenSSF 7.2 3.7K dependents Pushed 6d ago 4% patched ~1372d to patch Full package profile →

Do you use tensorflow? You're affected.

Severity & Risk

CVSS 3.1

6.5 / 10

EPSS

0.3%

chance of exploitation in 30 days

Higher than 52% of all CVEs

Source: EPSS v3 — FIRST.org

Exploitation Status

Exploit Available

Exploitation: MEDIUM

Sophistication

Moderate

Exploitation Confidence

medium

○ CISA SSVC: Public PoC

○ Public PoC indexed (trickest/cve)

Composite signal derived from CISA KEV, CISA SSVC, EPSS, trickest/cve, and Nuclei templates.

Attack Surface

AV Network

AC Low

PR Low

UI None

S Unchanged

C None

I None

A High

Recommended Action

6 steps

Patch: upgrade to TensorFlow 2.8.0, or cherry-pick to 2.7.1, 2.6.3, or 2.5.3.
Validate SavedModel integrity at ingestion: verify SHA-256 checksums and digital signatures before loading.
Restrict model upload permissions — only allow trusted CI/CD pipelines or verified model registries to push SavedModels to serving infrastructure.
Run model loading in isolated sandboxed processes with resource limits so a crash does not take down the entire inference service.
Monitor for unexpected process exits or CHECK failure logs in TensorFlow serving logs as an indicator of exploitation attempts.
Enforce protobuf schema validation on SavedModel files before deserialization.

CISA SSVC Assessment

Decision Track*

Exploitation poc

Automatable No

Technical Impact partial

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

Classification

DoS Supply Chain Framework Model Inference AML.T0010.003 - Model AML.T0011.000 - Unsafe AI Artifacts AML.T0018 - Manipulate AI Model AML.T0029 - Denial of AI Service AML.T0043.003 - Manual Modification

Compliance Impact

This CVE is relevant to:

EU AI Act

Article 15 - Accuracy, Robustness and Cybersecurity Article 9 - Risk Management System

ISO 42001

A.6.2 - AI System Design and Development — System Robustness A.9.3 - AI System Operation — Integrity of AI Inputs

NIST AI RMF

GOVERN 1.1 - AI Risk Policies and Accountability MANAGE 2.2 - Mechanisms to Sustain Value of Deployed AI Systems

OWASP LLM Top 10

LLM05:2025 - Improper Output Handling / Supply Chain Vulnerabilities

Frequently Asked Questions

What is CVE-2022-23583?

A low-privileged attacker can crash TensorFlow inference services by supplying a malicious SavedModel with mismatched dtype fields, causing CHECK assertion failures. Any platform that loads externally-supplied or user-provided SavedModels is directly exposed. Patch to TF 2.8.0, 2.7.1, 2.6.3, or 2.5.3 immediately and restrict who can push models to your serving infrastructure.

Is CVE-2022-23583 actively exploited?

Proof-of-concept exploit code is publicly available for CVE-2022-23583, increasing the risk of exploitation.

How to fix CVE-2022-23583?

1. Patch: upgrade to TensorFlow 2.8.0, or cherry-pick to 2.7.1, 2.6.3, or 2.5.3. 2. Validate SavedModel integrity at ingestion: verify SHA-256 checksums and digital signatures before loading. 3. Restrict model upload permissions — only allow trusted CI/CD pipelines or verified model registries to push SavedModels to serving infrastructure. 4. Run model loading in isolated sandboxed processes with resource limits so a crash does not take down the entire inference service. 5. Monitor for unexpected process exits or CHECK failure logs in TensorFlow serving logs as an indicator of exploitation attempts. 6. Enforce protobuf schema validation on SavedModel files before deserialization.

What systems are affected by CVE-2022-23583?

This vulnerability affects the following AI/ML architecture patterns: model serving, training pipelines, multi-tenant ML platforms, model registries, inference endpoints.

What is the CVSS score for CVE-2022-23583?

CVE-2022-23583 has a CVSS v3.1 base score of 6.5 (MEDIUM). The EPSS exploitation probability is 0.29%.

Technical Details

NVD Description

Tensorflow is an Open Source Machine Learning Framework. A malicious user can cause a denial of service by altering a `SavedModel` such that any binary op would trigger `CHECK` failures. This occurs when the protobuf part corresponding to the tensor arguments is modified such that the `dtype` no longer matches the `dtype` expected by the op. In that case, calling the templated binary operator for the binary op would receive corrupted data, due to the type confusion involved. If `Tin` and `Tout` don't match the type of data in `out` and `input_*` tensors then `flat<*>` would interpret it wrongly. In most cases, this would be a silent failure, but we have noticed scenarios where this results in a `CHECK` crash, hence a denial of service. The fix will be included in TensorFlow 2.8.0. We will also cherrypick this commit on TensorFlow 2.7.1, TensorFlow 2.6.3, and TensorFlow 2.5.3, as these are also affected and still in supported range.

Exploitation Scenario

An adversary with a low-privilege account on a shared ML platform (e.g., a researcher account on an internal Kubeflow or MLflow deployment) crafts a SavedModel where the protobuf tensor descriptors for a binary op have a deliberately mismatched dtype — e.g., declaring float32 but providing int32 data. They upload this model to the model registry. When the inference service loads and executes the model, the templated binary operator in cwise_ops_common.h receives corrupted type data, the flat<> accessor interprets memory with the wrong type, and TensorFlow triggers a CHECK failure that crashes the serving process. If the service lacks process isolation or auto-restart, all tenants on that inference endpoint experience a complete outage until an operator intervenes.