Memory Corruption in vLLM Audio Transcription

Executive Summary

The vulnerability exists in vLLM, an inference and serving engine for large language models, specifically in versions prior to 0.23.1rc0. The issue is with the /v1/audio/transcriptions endpoint, which limits the size of compressed audio uploads but does not limit the size of the decoded PCM output. For example, a 25MB OPUS audio file can expand to approximately 14.9GB of float32 PCM data when decoded, potentially causing resource exhaustion.

Useful 0 Not Useful 0

Impact Analysis

This vulnerability can lead to a denial of service (DoS) condition by exhausting system resources such as memory or processing power due to the large expansion of audio data during decoding. An attacker could exploit this by sending a relatively small compressed audio file that expands to a very large size, overwhelming the system and causing it to become unresponsive or crash.

Useful 0 Not Useful 0

Mitigation Strategies

To mitigate this vulnerability, upgrade vLLM to version 0.23.1rc0 or later, where the issue with the /v1/audio/transcriptions endpoint has been fixed.

Useful 0 Not Useful 0

Compliance Impact

The vulnerability in vLLM allows an attacker to cause a denial-of-service (DoS) condition by exhausting server memory through an audio decompression bomb. This impacts system availability, which is a key aspect of many compliance standards such as GDPR and HIPAA that require maintaining availability and integrity of systems processing sensitive data.

While the CVE description and resources do not explicitly mention compliance with specific standards like GDPR or HIPAA, the denial-of-service caused by this vulnerability could lead to service outages or interruptions, potentially violating availability requirements under these regulations.

No direct information is provided about data confidentiality or integrity impacts, so the main compliance concern relates to availability and service continuity.

Useful 0 Not Useful 0

Detection Guidance

This vulnerability can be detected by monitoring the /v1/audio/transcriptions endpoint for unusually high memory usage or out-of-memory (OOM) conditions triggered by audio uploads that are within the 25MB compressed size limit but decompress into very large PCM data.

To detect potential exploitation attempts, you can look for requests uploading audio files close to the 25MB limit and observe if the server memory usage spikes significantly during decoding.

Suggested commands include using network monitoring and system resource tools such as:

• Use tcpdump or Wireshark to capture and filter HTTP POST requests to the /v1/audio/transcriptions endpoint with payload sizes near 25MB.
• Use system monitoring commands like `top`, `htop`, or `free -m` to observe memory usage spikes during audio transcription requests.
• Use application logs to identify errors related to audio decoding failures or memory exhaustion.
• If you have access to the vLLM environment, check if the environment variable `VLLM_MAX_AUDIO_DECODE_DURATION_S` is set to limit decoded audio duration, which mitigates the vulnerability.

Useful 0 Not Useful 0