CVE-2025-12343: FFmpeg TensorFlow Backend DoS Vulnerability

CVE-2025-12343 Overview

A double-free vulnerability has been identified in FFmpeg's TensorFlow backend within the libavfilter/dnn_backend_tf.c source file. The flaw exists in the dnn_execute_model_tf() function, where a task object is freed multiple times in certain error-handling paths. This redundant memory deallocation can lead to a double-free condition, potentially causing FFmpeg or any application using it to crash when processing TensorFlow-based DNN models. This results in a denial-of-service scenario but does not allow arbitrary code execution under normal conditions.

Critical Impact

Applications utilizing FFmpeg with TensorFlow-based deep neural network (DNN) model processing may experience crashes or instability when encountering malformed inputs that trigger the vulnerable error-handling paths.

Affected Products

• FFmpeg with TensorFlow DNN backend enabled
• Applications and media processing pipelines utilizing FFmpeg's DNN filtering capabilities
• Systems processing untrusted media through FFmpeg's TensorFlow integration

Discovery Timeline

• 2026-02-18 - CVE-2025-12343 published to NVD
• 2026-02-19 - Last updated in NVD database

Technical Details for CVE-2025-12343

Vulnerability Analysis

This vulnerability is classified as CWE-415 (Double Free), a memory corruption flaw that occurs when a program attempts to free the same memory location more than once. In the context of FFmpeg's TensorFlow backend, the issue manifests within the dnn_execute_model_tf() function during error handling.

When processing TensorFlow-based DNN models, if certain error conditions are encountered, the code path may inadvertently call free operations on the same task object multiple times. This can corrupt the memory allocator's internal data structures, leading to undefined behavior. In most cases, this results in application crashes, causing a denial-of-service condition for media processing applications.

The vulnerability requires local access and user interaction to exploit, as an attacker would need to convince a user to process a specially crafted input through FFmpeg's TensorFlow DNN backend. While the immediate impact is limited to availability (denial of service), double-free vulnerabilities in some contexts can potentially be leveraged for more severe attacks, though this particular instance does not appear to allow arbitrary code execution under normal conditions.

Root Cause

The root cause is improper memory management in the error-handling logic within libavfilter/dnn_backend_tf.c. Specifically, the dnn_execute_model_tf() function contains code paths where a task object's memory is deallocated multiple times when errors occur during TensorFlow model execution. This indicates missing or incorrect cleanup state tracking that would normally prevent redundant free operations.

Attack Vector

The attack vector requires local access to the system where FFmpeg is running. An attacker must craft a malicious input (such as a specially constructed media file or DNN model) that, when processed by FFmpeg's TensorFlow backend, triggers the error-handling paths containing the double-free condition. User interaction is required to initiate the processing of the malicious input.

The exploitation scenario involves:

1. Attacker creates a malformed input designed to trigger specific error conditions
2. User processes the input through FFmpeg with TensorFlow DNN filtering enabled
3. The vulnerable error-handling code executes, causing the double-free
4. FFmpeg crashes, denying service to the media processing pipeline

The vulnerability mechanism involves improper memory deallocation in error handling. When the dnn_execute_model_tf() function encounters certain error conditions, the cleanup code may execute multiple times on the same task object, leading to the double-free condition. For detailed technical information, see the Red Hat Bug Report #2406533.

Detection Methods for CVE-2025-12343

Indicators of Compromise

• Unexpected FFmpeg crashes or segmentation faults during DNN model processing
• Core dumps containing stack traces pointing to dnn_execute_model_tf() or related TensorFlow backend functions
• Memory corruption errors in system logs when FFmpeg processes certain media files

Detection Strategies

• Monitor for FFmpeg process crashes, particularly when TensorFlow DNN filtering is in use
• Implement memory error detection tools (such as AddressSanitizer) during testing phases to identify double-free conditions
• Review application logs for repeated failures during DNN model execution
• Deploy runtime memory protection mechanisms that can detect and alert on double-free attempts

Monitoring Recommendations

• Enable detailed logging for FFmpeg operations, particularly for DNN backend activities
• Set up alerting for unexpected FFmpeg process terminations in production environments
• Monitor system stability metrics for applications relying on FFmpeg's TensorFlow integration
• Consider implementing crash reporting mechanisms to capture and analyze FFmpeg failures

How to Mitigate CVE-2025-12343

Immediate Actions Required

• Review your FFmpeg deployment to determine if TensorFlow DNN backend functionality is enabled and in use
• Limit processing of untrusted media files through FFmpeg's TensorFlow DNN filters where possible
• Implement input validation and sandboxing for media processing pipelines
• Monitor for vendor security patches addressing this vulnerability

Patch Information

At the time of publication, administrators should monitor the Red Hat CVE Advisory for patch availability and updates. The Red Hat Bug Report #2406533 provides additional technical details and tracking information for this vulnerability.

Workarounds

• Disable TensorFlow DNN backend functionality in FFmpeg if not required for your use case
• Process media files from trusted sources only when TensorFlow DNN filtering is necessary
• Implement process isolation and sandboxing for FFmpeg operations to limit crash impact
• Configure resource limits and automatic restart policies for FFmpeg processes to maintain availability

# Configuration example
# Verify if FFmpeg is compiled with TensorFlow support
ffmpeg -buildconf 2>&1 | grep -i tensorflow

# If TensorFlow DNN backend is not needed, consider using FFmpeg builds without TensorFlow support
# Or explicitly avoid using DNN filters that require TensorFlow backend
# Example: Avoid filters like dnn_processing with tensorflow backend