CVE-2021-29519: TensorFlow type confusion DoS

CISO Take

A type confusion bug in TensorFlow's SparseCross op lets any local low-privilege user crash TF-based workloads via a CHECK-failure abort — no data exposure, availability impact only. Patch to TF 2.5.0+ or apply the backports (2.4.2, 2.3.3, 2.2.3, 2.1.4). Priority is moderate unless you run shared ML infrastructure where untrusted users can submit training jobs or model inputs.

What is the risk?

Medium risk in practice. Local attack vector with low-privilege requirement significantly limits exposure — a remote attacker needs a foothold first, or the vulnerable op must be reachable via a model serving endpoint processing user-supplied sparse features. No confidentiality or integrity impact (C:N/I:N/A:H). Not in CISA KEV, no evidence of active exploitation. Risk elevates meaningfully in shared GPU clusters, multi-tenant Jupyter/MLflow environments, or online feature preprocessing services that invoke SparseCross on user-controlled categorical inputs.

What systems are affected?

Package	Ecosystem	Vulnerable Range	Patched
TensorFlow	pip	—	No patch
195.8K OpenSSF 7.1 3.7K dependents Pushed 2d ago 4% patched ~1372d to patch Full package profile →

Do you use TensorFlow? You're affected.

How severe is it?

CVSS 3.1

5.5 / 10

EPSS

0.2%

chance of exploitation in 30 days

Higher than 9% of all CVEs

Source: EPSS v3 — FIRST.org

Exploitation Status

Exploit Available

Exploitation: MEDIUM

Sophistication

Trivial

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?

AV Local

AC Low

PR Low

UI None

S Unchanged

C None

I None

A High

What should I do?

5 steps

PATCH

Upgrade to TF 2.5.0+. Backports available for TF 2.4.2, 2.3.3, 2.2.3, 2.1.4 — apply the minimum backport matching your installed branch. Commit: b1cc5e5a50e7cee09f2c6eb48eb40ee9c4125025.
INPUT VALIDATION

Add explicit dtype guards before SparseCross calls — reject inputs that mix DT_STRING and DT_INT64 types at the application layer.
ISOLATION

Ensure TF Serving instances handling external inputs run in sandboxed containers with restart policies; limit blast radius of process crashes.
DETECTION

Alert on repeated TF process SIGABRT/CHECK failure signals (look for 'Check failed' in TF logs); anomalous crash patterns from the same user or endpoint warrant investigation.
INVENTORY

Audit all TF versions across ML training, serving, and preprocessing workloads — unpatched versions in containerized ML pipelines are the highest-priority targets.

How is it classified?

DoS Framework AML.T0010.001 - AI Software AML.T0029 - Denial of AI Service AML.T0043.003 - Manual Modification

Which compliance frameworks are affected?

This CVE is relevant to:

EU AI Act

Article 15 - Accuracy, robustness and cybersecurity

ISO 42001

A.6.1.6 - AI system security

NIST AI RMF

MANAGE 2.2 - Mechanisms to sustain AI risk management

OWASP LLM Top 10

LLM03 - Supply Chain Vulnerabilities

Frequently Asked Questions

What is CVE-2021-29519?

A type confusion bug in TensorFlow's SparseCross op lets any local low-privilege user crash TF-based workloads via a CHECK-failure abort — no data exposure, availability impact only. Patch to TF 2.5.0+ or apply the backports (2.4.2, 2.3.3, 2.2.3, 2.1.4). Priority is moderate unless you run shared ML infrastructure where untrusted users can submit training jobs or model inputs.

Is CVE-2021-29519 actively exploited?

Proof-of-concept exploit code is publicly available for CVE-2021-29519, increasing the risk of exploitation.

How to fix CVE-2021-29519?

1. PATCH: Upgrade to TF 2.5.0+. Backports available for TF 2.4.2, 2.3.3, 2.2.3, 2.1.4 — apply the minimum backport matching your installed branch. Commit: b1cc5e5a50e7cee09f2c6eb48eb40ee9c4125025. 2. INPUT VALIDATION: Add explicit dtype guards before SparseCross calls — reject inputs that mix DT_STRING and DT_INT64 types at the application layer. 3. ISOLATION: Ensure TF Serving instances handling external inputs run in sandboxed containers with restart policies; limit blast radius of process crashes. 4. DETECTION: Alert on repeated TF process SIGABRT/CHECK failure signals (look for 'Check failed' in TF logs); anomalous crash patterns from the same user or endpoint warrant investigation. 5. INVENTORY: Audit all TF versions across ML training, serving, and preprocessing workloads — unpatched versions in containerized ML pipelines are the highest-priority targets.

What systems are affected by CVE-2021-29519?

This vulnerability affects the following AI/ML architecture patterns: training pipelines, model serving, feature engineering pipelines.

What is the CVSS score for CVE-2021-29519?

CVE-2021-29519 has a CVSS v3.1 base score of 5.5 (MEDIUM). The EPSS exploitation probability is 0.19%.

What is the AI security impact?

Affected AI Architectures

training pipelinesmodel servingfeature engineering pipelines

MITRE ATLAS Techniques

AML.T0010.001 AI Software

AML.T0029 Denial of AI Service

AML.T0043.003 Manual Modification

Compliance Controls Affected

EU AI Act: Article 15

ISO 42001: A.6.1.6

NIST AI RMF: MANAGE 2.2

OWASP LLM Top 10: LLM03

What are the technical details?

Original Advisory

TensorFlow is an end-to-end open source platform for machine learning. The API of `tf.raw_ops.SparseCross` allows combinations which would result in a `CHECK`-failure and denial of service. This is because the implementation(https://github.com/tensorflow/tensorflow/blob/3d782b7d47b1bf2ed32bd4a246d6d6cadc4c903d/tensorflow/core/kernels/sparse_cross_op.cc#L114-L116) is tricked to consider a tensor of type `tstring` which in fact contains integral elements. Fixing the type confusion by preventing mixing `DT_STRING` and `DT_INT64` types solves this issue. The fix will be included in TensorFlow 2.5.0. We will also cherrypick this commit on TensorFlow 2.4.2, TensorFlow 2.3.3, TensorFlow 2.2.3 and TensorFlow 2.1.4, as these are also affected and still in supported range.

Exploitation Scenario

An adversary with access to a shared MLOps environment — a data scientist account on a multi-tenant JupyterHub, an MLflow experiment runner, or a compromised CI/CD pipeline — writes a short Python snippet invoking tf.raw_ops.SparseCross and passes it a tstring-typed tensor populated with integer values. When the op executes, TF's internal CHECK fires and aborts the process. In a co-located GPU cluster, this terminates other users' training runs. A more impactful scenario: an online serving endpoint preprocessing user-supplied sparse categorical features (e.g., for a recommendation model) uses SparseCross in the feature pipeline. An external attacker crafts HTTP requests with malformed feature payloads containing mixed-type tensors, repeatedly crashing the serving process and causing sustained service unavailability for all users of that model.

Weaknesses (CWE)

CWE-843 Access of Resource Using Incompatible Type ('Type Confusion')

CWE-843 — Access of Resource Using Incompatible Type ('Type Confusion'): The product allocates or initializes a resource such as a pointer, object, or variable using one type, but it later accesses that resource using a type that is incompatible with the original type.

Source: MITRE CWE corpus.