Executive Summary

CVE-2026-56340 is a vulnerability in vLLM versions 0.10.2 up to but not including 0.13.0, where sparse tensor validation is missing in the multimodal embeddings processing feature called prompt-embeds.

Because PyTorch disables sparse tensor invariant checks by default, attackers can submit specially crafted embedding requests with malformed tensor indices (such as negative or out-of-bounds values). This can cause crashes, resource exhaustion (denial of service), or even memory corruption including out-of-bounds or write-what-where conditions.

This vulnerability is a continuation of CVE-2025-62164, where the previous fix only disabled the feature by default rather than fixing the root cause.

Useful 0 Not Useful 0

Impact Analysis

This vulnerability can impact you by allowing an attacker to cause denial of service (DoS) through crashes or resource exhaustion when the prompt-embeds feature is enabled.

There is also potential for memory corruption vulnerabilities, such as out-of-bounds writes or write-what-where conditions, which could lead to remote code execution or other serious security breaches.

The attack requires low privileges and no user interaction, making it easier to exploit remotely.

Useful 0 Not Useful 0

Mitigation Strategies

To mitigate this vulnerability, you should disable the prompt-embeds feature in vLLM if it is currently enabled, as this feature is the attack vector for malformed sparse tensor indices.

Additionally, upgrade vLLM to a version later than 0.13.0 where the root cause is addressed by adding proper sparse tensor validation to ensure indices are valid, non-negative, and within bounds.

Until the fix is applied, avoid enabling the prompt-embeds feature to prevent denial-of-service or potential memory corruption attacks.

Useful 0 Not Useful 0