CVE-2021-37669 — MEDIUM (CVSS 5.5) AI Security Vulnerability

CISO Take

This is a local denial-of-service in TensorFlow's object detection operations (NonMaxSuppression, CombinedNonMaxSuppression) caused by a negative integer being implicitly cast to unsigned, crashing the process. If your model serving endpoints expose these ops to user-controlled inputs, an attacker can crash your inference service with a single malformed request. Patch to TF 2.6.0 or the backport releases (2.5.1, 2.4.3, 2.3.4) immediately.

Risk Assessment

Medium risk in isolation, higher in model serving contexts. CVSS 5.5 (Local) understates real-world exposure: model serving APIs accepting arbitrary inference requests effectively lower the attack vector to Network. No data exfiltration or code execution possible — pure availability impact. Exploitation requires no AI/ML knowledge, just a negative integer in the right field. Not in CISA KEV and no known active exploitation, but the technique is trivially reproducible from the public advisory.

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 9% of all CVEs

Source: EPSS v3 — FIRST.org

Exploitation Status

No known exploitation

Sophistication

Trivial

Attack Surface

AV Local

AC Low

PR Low

UI None

S Unchanged

C None

I None

A High

Recommended Action

5 steps

PATCH

Upgrade to TensorFlow 2.6.0+. If pinned, apply cherry-picks: commit 3a7362750d5c (NonMaxSuppression fix) and b5cdbf12ffca (CombinedNonMaxSuppression fix).
INPUT VALIDATION

Add server-side validation rejecting non-positive values for output_size / max_output_size parameters before they reach TF ops.
DETECTION

Monitor inference service crash rates and restarts; sudden spikes on endpoints accepting detection models indicate exploitation attempts.
SANDBOXING

Run TF Serving instances in containers with restart policies so DoS impact is limited to brief availability windows.
AUDIT

Inventory all internal services using tf.raw_ops.NonMaxSuppressionV5 or CombinedNonMaxSuppression directly.

Classification

DoS Framework Inference AML.T0029 - Denial of AI Service AML.T0034 - Cost Harvesting AML.T0049 - Exploit Public-Facing Application

Compliance Impact

This CVE is relevant to:

EU AI Act

Article 15 - Accuracy, robustness and cybersecurity

ISO 42001

A.6.1.2 - AI risk assessment A.6.2.4 - AI system operation and monitoring

NIST AI RMF

GOVERN 1.7 - Processes for AI risk monitoring MEASURE 2.5 - AI system availability and resiliency

OWASP LLM Top 10

LLM04 - Model Denial of Service

Frequently Asked Questions

What is CVE-2021-37669?

This is a local denial-of-service in TensorFlow's object detection operations (NonMaxSuppression, CombinedNonMaxSuppression) caused by a negative integer being implicitly cast to unsigned, crashing the process. If your model serving endpoints expose these ops to user-controlled inputs, an attacker can crash your inference service with a single malformed request. Patch to TF 2.6.0 or the backport releases (2.5.1, 2.4.3, 2.3.4) immediately.

Is CVE-2021-37669 actively exploited?

No confirmed active exploitation of CVE-2021-37669 has been reported, but organizations should still patch proactively.

How to fix CVE-2021-37669?

1. PATCH: Upgrade to TensorFlow 2.6.0+. If pinned, apply cherry-picks: commit 3a7362750d5c (NonMaxSuppression fix) and b5cdbf12ffca (CombinedNonMaxSuppression fix). 2. INPUT VALIDATION: Add server-side validation rejecting non-positive values for output_size / max_output_size parameters before they reach TF ops. 3. DETECTION: Monitor inference service crash rates and restarts; sudden spikes on endpoints accepting detection models indicate exploitation attempts. 4. SANDBOXING: Run TF Serving instances in containers with restart policies so DoS impact is limited to brief availability windows. 5. AUDIT: Inventory all internal services using tf.raw_ops.NonMaxSuppressionV5 or CombinedNonMaxSuppression directly.

What systems are affected by CVE-2021-37669?

This vulnerability affects the following AI/ML architecture patterns: model serving, inference pipelines, object detection pipelines, training pipelines.

What is the CVSS score for CVE-2021-37669?

CVE-2021-37669 has a CVSS v3.1 base score of 5.5 (MEDIUM). The EPSS exploitation probability is 0.03%.

Technical Details

NVD Description

TensorFlow is an end-to-end open source platform for machine learning. In affected versions an attacker can cause denial of service in applications serving models using `tf.raw_ops.NonMaxSuppressionV5` by triggering a division by 0. The [implementation](https://github.com/tensorflow/tensorflow/blob/460e000de3a83278fb00b61a16d161b1964f15f4/tensorflow/core/kernels/image/non_max_suppression_op.cc#L170-L271) uses a user controlled argument to resize a `std::vector`. However, as `std::vector::resize` takes the size argument as a `size_t` and `output_size` is an `int`, there is an implicit conversion to unsigned. If the attacker supplies a negative value, this conversion results in a crash. A similar issue occurs in `CombinedNonMaxSuppression`. We have patched the issue in GitHub commit 3a7362750d5c372420aa8f0caf7bf5b5c3d0f52d and commit [b5cdbf12ffcaaffecf98f22a6be5a64bb96e4f58. The fix will be included in TensorFlow 2.6.0. We will also cherrypick this commit on TensorFlow 2.5.1, TensorFlow 2.4.3, and TensorFlow 2.3.4, as these are also affected and still in supported range.

Exploitation Scenario

An attacker probes a public-facing object detection API (e.g., a retail image tagging service or autonomous vehicle inference endpoint powered by TF Serving). They craft a POST request to the inference endpoint with max_detections set to -1 or another negative integer. TensorFlow's NonMaxSuppression kernel receives the value, implicitly converts it to size_t (becoming a massive unsigned integer), attempts to allocate ~18 exabytes of memory, and crashes. With no restart policy, the service goes down. With a simple restart policy, the attacker can loop requests to maintain a continuous DoS at negligible cost. This can be scripted in under 10 lines of Python using the TF Serving REST API.