CVE-2022-23559 — HIGH (CVSS 8.8) AI Security Vulnerability

CISO Take

Any service that accepts user-submitted TFLite model files is exposed to heap out-of-bounds read/write, which can lead to remote code execution with only low-privilege access and no user interaction required. If your organisation runs TFLite inference endpoints or pipelines that ingest externally supplied models, patch immediately and enforce strict model provenance controls. CVSS 8.8 with network vector and low complexity makes this trivially weaponisable by any authenticated API user.

Risk Assessment

High severity (CVSS 8.8). The attack requires only low privileges and zero user interaction, eliminating most practical barriers to exploitation. The integer overflow in embedding_size × lookup_size produces a heap OOB R/W primitive—a reliable stepping stone to RCE on the host running inference. Exposure is highest in multi-tenant model-serving platforms, ML-as-a-Service offerings, and CI/CD pipelines that automatically evaluate community-contributed models. Not in CISA KEV as of analysis date, but the low exploitation complexity warrants treatment as actively exploitable.

Affected Systems

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

Do you use tensorflow? You're affected.

Severity & Risk

CVSS 3.1

8.8 / 10

EPSS

0.5%

chance of exploitation in 30 days

Higher than 67% 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 High

I High

A High

Recommended Action

5 steps

PATCH

Upgrade TensorFlow to a version containing commits 1de4972, a4e401d, and f19be71 (released post-2022-02-04). Verify via pip show tensorflow and cross-check against TF security advisories.
VALIDATE INPUTS

Before passing a TFLite model to the interpreter, validate that embedding_size and lookup_size values are within safe bounds and their product does not overflow int32/int64.
MODEL PROVENANCE

Reject TFLite models from untrusted sources; enforce cryptographic signing of model artifacts in your ML pipeline.
SANDBOX INFERENCE

Run TFLite inference in isolated processes or containers with minimal filesystem and network permissions; use seccomp/AppArmor profiles.
DETECT

Monitor inference workers for anomalous memory access patterns (ASAN in staging, heap-hardening allocators in production). Alert on unexpected crashes or segfaults in TFLite interpreter processes.

CISA SSVC Assessment

Decision Attend

Exploitation poc

Automatable No

Technical Impact total

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

Classification

Supply Chain Code Execution Framework Inference AML.T0010.001 - AI Software AML.T0011.000 - Unsafe AI Artifacts AML.T0049 - Exploit Public-Facing Application

Compliance Impact

This CVE is relevant to:

EU AI Act

Article 15 - Accuracy, robustness and cybersecurity Article 9 - Risk management system

ISO 42001

A.9.3 - AI system security

NIST AI RMF

MANAGE-2.2 - Risk treatments applied to deployed AI systems

OWASP LLM Top 10

LLM05:2025 - Supply Chain Vulnerabilities

Frequently Asked Questions

What is CVE-2022-23559?

Any service that accepts user-submitted TFLite model files is exposed to heap out-of-bounds read/write, which can lead to remote code execution with only low-privilege access and no user interaction required. If your organisation runs TFLite inference endpoints or pipelines that ingest externally supplied models, patch immediately and enforce strict model provenance controls. CVSS 8.8 with network vector and low complexity makes this trivially weaponisable by any authenticated API user.

Is CVE-2022-23559 actively exploited?

Proof-of-concept exploit code is publicly available for CVE-2022-23559, increasing the risk of exploitation.

How to fix CVE-2022-23559?

1. PATCH: Upgrade TensorFlow to a version containing commits 1de4972, a4e401d, and f19be71 (released post-2022-02-04). Verify via `pip show tensorflow` and cross-check against TF security advisories. 2. VALIDATE INPUTS: Before passing a TFLite model to the interpreter, validate that embedding_size and lookup_size values are within safe bounds and their product does not overflow int32/int64. 3. MODEL PROVENANCE: Reject TFLite models from untrusted sources; enforce cryptographic signing of model artifacts in your ML pipeline. 4. SANDBOX INFERENCE: Run TFLite inference in isolated processes or containers with minimal filesystem and network permissions; use seccomp/AppArmor profiles. 5. DETECT: Monitor inference workers for anomalous memory access patterns (ASAN in staging, heap-hardening allocators in production). Alert on unexpected crashes or segfaults in TFLite interpreter processes.

What systems are affected by CVE-2022-23559?

This vulnerability affects the following AI/ML architecture patterns: TFLite edge inference, model serving, training pipelines, recommendation system inference, on-device ML (mobile/embedded).

What is the CVSS score for CVE-2022-23559?

CVE-2022-23559 has a CVSS v3.1 base score of 8.8 (HIGH). The EPSS exploitation probability is 0.52%.

Technical Details

NVD Description

Tensorflow is an Open Source Machine Learning Framework. An attacker can craft a TFLite model that would cause an integer overflow in embedding lookup operations. Both `embedding_size` and `lookup_size` are products of values provided by the user. Hence, a malicious user could trigger overflows in the multiplication. In certain scenarios, this can then result in heap OOB read/write. Users are advised to upgrade to a patched version.

Exploitation Scenario

An adversary with low-privilege access to an internal ML model evaluation service uploads a crafted TFLite model. The model contains a sparse embedding layer with manipulated metadata: embedding_size and lookup_size are chosen such that their product overflows a 32-bit integer, yielding a small allocation. Subsequent kernel operations write attacker-controlled data beyond the allocation boundary. With a suitable heap layout (achievable through multiple requests), the adversary overwrites adjacent allocator metadata or function pointers, redirecting execution to a shellcode stub or ROP chain. The inference worker process is compromised, granting access to model weights, inference request data, and potentially the host if container hardening is absent.