CVE-2020-15201: TensorFlow heap overflow

Q: Is CVE-2020-15201 actively exploited?

Proof-of-concept exploit code is publicly available for CVE-2020-15201, increasing the risk of exploitation.

Q: How to fix CVE-2020-15201?

1. Patch: Upgrade TensorFlow to 2.3.1 or later — patch is in commit 3cbb917b4714766030b28eba9fb41bb97ce9ee02. 2. Validate inputs: Add server-side validation that splits tensors form valid ragged partitions before invoking RaggedCountSparseOutput; reject any request where splits[-1] < num_values. 3. Network segmentation: TensorFlow Serving gRPC/REST endpoints should never be directly internet-exposed; place behind an API gateway that validates tensor shapes and types. 4. Detection: Monitor for malformed tensor shape errors or unexpected crashes in TF Serving logs; anomalous spike in op-level errors on sparse/ragged ops may indicate probing. 5. Runtime hardening: Run inference containers with restricted memory access (seccomp, read-only rootfs) to limit blast radius of memory corruption.

Q: What systems are affected by CVE-2020-15201?

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

Q: What is the CVSS score for CVE-2020-15201?

CVE-2020-15201 has a CVSS v3.1 base score of 4.8 (MEDIUM). The EPSS exploitation probability is 0.56%.

CISO Take

A heap buffer overflow in TensorFlow's RaggedCountSparseOutput operation allows network-accessible exploitation with no authentication required, though high attack complexity limits immediate risk. Any TensorFlow deployment exposing inference endpoints that process ragged/sparse tensor inputs should patch to 2.3.1+ immediately. No active exploitation reported, but the no-auth network vector is unacceptable for production AI serving infrastructure.

What is the risk?

Medium risk overall, but context-dependent severity for AI serving deployments. CVSS 4.8 reflects high attack complexity (AC:H), which requires crafting malformed ragged tensor inputs with precisely invalid splits partitioning. The network vector (AV:N) with no privileges (PR:N) and no user interaction (UI:N) is the primary risk amplifier — any internet-exposed TensorFlow Serving or gRPC inference endpoint accepting raw tensor inputs is a legitimate attack surface. No CISA KEV listing and 2020 publication date suggest no widespread active exploitation.

What systems are affected?

Package	Ecosystem	Vulnerable Range	Patched
TensorFlow	pip	—	No patch
195.8K OpenSSF 7.1 3.7K dependents Pushed 5d ago 4% patched ~1372d to patch Full package profile →

Do you use TensorFlow? You're affected.

How severe is it?

CVSS 3.1

4.8 / 10

EPSS

0.6%

chance of exploitation in 30 days

Higher than 43% of all CVEs

Source: EPSS v3 — FIRST.org

Exploitation Status

Exploit Available

Exploitation: MEDIUM

Sophistication

Moderate

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 Network

AC High

PR None

UI None

S Unchanged

C Low

I Low

A None

What should I do?

5 steps

Patch: Upgrade TensorFlow to 2.3.1 or later — patch is in commit 3cbb917b4714766030b28eba9fb41bb97ce9ee02.
Validate inputs: Add server-side validation that splits tensors form valid ragged partitions before invoking RaggedCountSparseOutput; reject any request where splits[-1] < num_values.
Network segmentation: TensorFlow Serving gRPC/REST endpoints should never be directly internet-exposed; place behind an API gateway that validates tensor shapes and types.
Detection: Monitor for malformed tensor shape errors or unexpected crashes in TF Serving logs; anomalous spike in op-level errors on sparse/ragged ops may indicate probing.
Runtime hardening: Run inference containers with restricted memory access (seccomp, read-only rootfs) to limit blast radius of memory corruption.

How is it classified?

Code Execution Data Extraction Framework Inference AML.T0010.001 - AI Software AML.T0024 - Exfiltration via AI Inference API AML.T0049 - Exploit Public-Facing Application

Which compliance frameworks are affected?

This CVE is relevant to:

EU AI Act

Art.15 - Accuracy, robustness and cybersecurity

ISO 42001

8.4 - AI system risk management 9.1 - Monitoring, measurement, analysis and evaluation

NIST AI RMF

GOVERN-4.1 - Organizational teams are committed to transparency MANAGE-2.2 - Mechanisms to sustain management of AI risks

OWASP LLM Top 10

LLM09 - Overreliance (misnomer here — maps to insecure plugin/dependency)

Frequently Asked Questions

What is CVE-2020-15201?

A heap buffer overflow in TensorFlow's RaggedCountSparseOutput operation allows network-accessible exploitation with no authentication required, though high attack complexity limits immediate risk. Any TensorFlow deployment exposing inference endpoints that process ragged/sparse tensor inputs should patch to 2.3.1+ immediately. No active exploitation reported, but the no-auth network vector is unacceptable for production AI serving infrastructure.

Is CVE-2020-15201 actively exploited?

Proof-of-concept exploit code is publicly available for CVE-2020-15201, increasing the risk of exploitation.

How to fix CVE-2020-15201?

1. Patch: Upgrade TensorFlow to 2.3.1 or later — patch is in commit 3cbb917b4714766030b28eba9fb41bb97ce9ee02. 2. Validate inputs: Add server-side validation that splits tensors form valid ragged partitions before invoking RaggedCountSparseOutput; reject any request where splits[-1] < num_values. 3. Network segmentation: TensorFlow Serving gRPC/REST endpoints should never be directly internet-exposed; place behind an API gateway that validates tensor shapes and types. 4. Detection: Monitor for malformed tensor shape errors or unexpected crashes in TF Serving logs; anomalous spike in op-level errors on sparse/ragged ops may indicate probing. 5. Runtime hardening: Run inference containers with restricted memory access (seccomp, read-only rootfs) to limit blast radius of memory corruption.

What systems are affected by CVE-2020-15201?

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

What is the CVSS score for CVE-2020-15201?

CVE-2020-15201 has a CVSS v3.1 base score of 4.8 (MEDIUM). The EPSS exploitation probability is 0.56%.

What is the AI security impact?

Affected AI Architectures

model servingtraining pipelinesdata preprocessing pipelines

MITRE ATLAS Techniques

AML.T0010.001 AI Software

AML.T0024 Exfiltration via AI Inference API

AML.T0049 Exploit Public-Facing Application

Compliance Controls Affected

EU AI Act: Art.15

ISO 42001: 8.4, 9.1

NIST AI RMF: GOVERN-4.1, MANAGE-2.2

OWASP LLM Top 10: LLM09

What are the technical details?

Original Advisory

In Tensorflow before version 2.3.1, the `RaggedCountSparseOutput` implementation does not validate that the input arguments form a valid ragged tensor. In particular, there is no validation that the values in the `splits` tensor generate a valid partitioning of the `values` tensor. Hence, the code is prone to heap buffer overflow. If `split_values` does not end with a value at least `num_values` then the `while` loop condition will trigger a read outside of the bounds of `split_values` once `batch_idx` grows too large. The issue is patched in commit 3cbb917b4714766030b28eba9fb41bb97ce9ee02 and is released in TensorFlow version 2.3.1.

Exploitation Scenario

An adversary targeting an organization running TensorFlow 2.x inference services constructs a malicious prediction request containing a ragged tensor where the splits array does not correctly partition the values array — specifically, the final split value is less than the total number of values. When TensorFlow processes this input through RaggedCountSparseOutput, the while loop increments batch_idx beyond the bounds of split_values, triggering a heap buffer over-read. In a model serving context (e.g., TF Serving gRPC endpoint processing NLP inputs or sparse feature vectors), this could leak adjacent heap memory contents — potentially exposing cached inference results, model weight fragments, or internal runtime metadata. An attacker could iterate over multiple crafted payloads to incrementally reconstruct sensitive in-memory data.

Weaknesses (CWE)

CWE-787 Out-of-bounds Write Primary CWE-122 Heap-based Buffer Overflow CWE-20 Improper Input Validation

CWE-787 — Out-of-bounds Write: The product writes data past the end, or before the beginning, of the intended buffer.

[Requirements] Use a language that does not allow this weakness to occur or provides constructs that make this weakness easier to avoid. For example, many languages that perform their own memory management, such as Java and Perl, are not subject to buffer overflows. Other languages, such as Ada and C#, typically provide overflow protection, but the protection can be disabled by the programmer. Be wary that a language's interface to native code may still be subject to overflows, even if the language itself is theoretically safe.
[Architecture and Design] Use a vetted library or framework that does not allow this weakness to occur or provides constructs that make this weakness easier to avoid. Examples include the Safe C String Library (SafeStr) by Messier and Viega [REF-57], and the Strsafe.h library from Microsoft [REF-56]. These libraries provide safer versions of overflow-prone string-handling functions.

Source: MITRE CWE corpus.