CVE-2021-29550 — MEDIUM (CVSS 5.5) AI Security Vulnerability

CISO Take

A locally-exploitable divide-by-zero in TensorFlow's FractionalAvgPool op crashes the TF runtime when attacker-controlled tensor shapes cause output_size to reach zero. Exposure is limited to systems where untrusted users can submit TensorFlow operations (e.g., shared ML platforms, multi-tenant training environments). Patch immediately to TF 2.5.0 or the applicable backport (2.4.2, 2.3.3, 2.2.3, 2.1.4) and restrict who can submit raw TF ops.

Risk Assessment

Effective risk is LOW-MEDIUM for most organizations. The local attack vector (AV:L) and low-privilege requirement mean this is not remotely exploitable without first gaining system access. In multi-tenant ML platforms or Jupyter-style environments where users can execute arbitrary TF ops, risk elevates to MEDIUM since a single malicious notebook cell can crash a shared inference or training node. No active exploitation reported; not in CISA KEV. CVSS 5.5 is accurate for isolated deployments but underestimates impact in shared-compute AI infrastructure.

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

5.5 / 10

EPSS

0.0%

chance of exploitation in 30 days

Higher than 1% 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, CISA SSVC, EPSS, trickest/cve, and Nuclei templates.

Attack Surface

AV Local

AC Low

PR Low

UI None

S Unchanged

C None

I None

A High

Recommended Action

5 steps

Patch

Upgrade TensorFlow to 2.5.0 or backports 2.4.2 / 2.3.3 / 2.2.3 / 2.1.4 per commit 548b5eaf23685d86f722233d8fbc21d0a4aecb96. Run pip show tensorflow to confirm version.
Validate inputs

Add pre-op assertion that input_size[i] >= pooling_ratio[i] before invoking FractionalAvgPool; reject or sanitize out-of-range values at API boundaries.
Restrict op access

If running a shared ML platform, use TF op allowlists or sandboxed execution (e.g., TF Serving with input schema validation) to prevent raw op invocation by untrusted users.
Detect

Monitor for unexpected TF process crashes or segfaults in training/serving logs — repeated crashes on FractionalAvgPool inputs may indicate probing.
Isolate jobs

Run multi-tenant training jobs in separate containers/processes so a crash in one job cannot affect others.

Classification

DoS Framework Inference AML.T0010.001 - AI Software AML.T0029 - Denial of AI Service AML.T0049 - Exploit Public-Facing Application

Compliance Impact

This CVE is relevant to:

EU AI Act

Article 15 - Accuracy, robustness and cybersecurity

ISO 42001

8.4 - AI system operation A.9.2 - AI system security

NIST AI RMF

GOVERN-6.1 - Policies and procedures are in place for AI risk management MANAGE-2.2 - Mechanisms are in place and applied to sustain the value of deployed AI systems

OWASP LLM Top 10

LLM07:2025 - System Prompt Leakage / Insecure Plugin Design

Frequently Asked Questions

What is CVE-2021-29550?

A locally-exploitable divide-by-zero in TensorFlow's FractionalAvgPool op crashes the TF runtime when attacker-controlled tensor shapes cause output_size to reach zero. Exposure is limited to systems where untrusted users can submit TensorFlow operations (e.g., shared ML platforms, multi-tenant training environments). Patch immediately to TF 2.5.0 or the applicable backport (2.4.2, 2.3.3, 2.2.3, 2.1.4) and restrict who can submit raw TF ops.

Is CVE-2021-29550 actively exploited?

Proof-of-concept exploit code is publicly available for CVE-2021-29550, increasing the risk of exploitation.

How to fix CVE-2021-29550?

1. **Patch**: Upgrade TensorFlow to 2.5.0 or backports 2.4.2 / 2.3.3 / 2.2.3 / 2.1.4 per commit 548b5eaf23685d86f722233d8fbc21d0a4aecb96. Run `pip show tensorflow` to confirm version. 2. **Validate inputs**: Add pre-op assertion that `input_size[i] >= pooling_ratio[i]` before invoking FractionalAvgPool; reject or sanitize out-of-range values at API boundaries. 3. **Restrict op access**: If running a shared ML platform, use TF op allowlists or sandboxed execution (e.g., TF Serving with input schema validation) to prevent raw op invocation by untrusted users. 4. **Detect**: Monitor for unexpected TF process crashes or segfaults in training/serving logs — repeated crashes on FractionalAvgPool inputs may indicate probing. 5. **Isolate jobs**: Run multi-tenant training jobs in separate containers/processes so a crash in one job cannot affect others.

What systems are affected by CVE-2021-29550?

This vulnerability affects the following AI/ML architecture patterns: training pipelines, model serving, shared ML platforms, inference servers.

What is the CVSS score for CVE-2021-29550?

CVE-2021-29550 has a CVSS v3.1 base score of 5.5 (MEDIUM). The EPSS exploitation probability is 0.01%.

Technical Details

NVD Description

TensorFlow is an end-to-end open source platform for machine learning. An attacker can cause a runtime division by zero error and denial of service in `tf.raw_ops.FractionalAvgPool`. This is because the implementation(https://github.com/tensorflow/tensorflow/blob/acc8ee69f5f46f92a3f1f11230f49c6ac266f10c/tensorflow/core/kernels/fractional_avg_pool_op.cc#L85-L89) computes a divisor quantity by dividing two user controlled values. The user controls the values of `input_size[i]` and `pooling_ratio_[i]` (via the `value.shape()` and `pooling_ratio` arguments). If the value in `input_size[i]` is smaller than the `pooling_ratio_[i]`, then the floor operation results in `output_size[i]` being 0. The `DCHECK_GT` line is a no-op outside of debug mode, so in released versions of TF this does not trigger. Later, these computed values are used as arguments(https://github.com/tensorflow/tensorflow/blob/acc8ee69f5f46f92a3f1f11230f49c6ac266f10c/tensorflow/core/kernels/fractional_avg_pool_op.cc#L96-L99) to `GeneratePoolingSequence`(https://github.com/tensorflow/tensorflow/blob/acc8ee69f5f46f92a3f1f11230f49c6ac266f10c/tensorflow/core/kernels/fractional_pool_common.cc#L100-L108). There, the first computation is a division in a modulo operation. Since `output_length` can be 0, this results in runtime crashing. 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 ML training platform (e.g., internal Jupyter Hub, MLflow Projects, or a model-as-a-service endpoint accepting custom model files) uploads or executes a model containing a FractionalAvgPool layer configured with `pooling_ratio > input_size` (e.g., input shape [1, 2, 2, 1] with pooling_ratio [1.0, 3.0, 3.0, 1.0]). When the operator executes, `output_size` computes to 0, the DCHECK is suppressed in release builds, and the subsequent modulo operation divides by zero — crashing the TF runtime. In a shared compute cluster this takes down the entire training node, causing a denial of service for all co-located jobs. An attacker could automate this to keep the platform continuously unavailable.