Denial-of-Service in vLLM via Unbounded Cache Salt Parameter

Executive Summary

CVE-2026-100647 is a denial-of-service (DoS) vulnerability in vLLM versions before 0.29.0. It involves the cache_salt parameter in OpenAI-compatible and Anthropic API endpoints lacking maximum length validation. Attackers can send HTTP requests with very large salt values (hundreds of megabytes), which are processed on the single EngineCore scheduler thread. This triggers expensive pickle serialization and SHA-256 hashing, stalling the thread and denying service to all concurrent requests.

Useful 0 Not Useful 0

Detection Guidance

Monitor for unusually large HTTP request payloads containing the cache_salt parameter. Check for requests with salt values exceeding typical sizes (e.g., >1MB). Use network monitoring tools like tcpdump or Wireshark to inspect HTTP request sizes and payloads for large cache_salt values.

Useful 0 Not Useful 0

Impact Analysis

This vulnerability allows unauthenticated remote attackers to exhaust CPU resources by sending specially crafted HTTP requests. It can cause severe delays or complete unavailability of the service for all users. For example, a 100 MB salt increases request latency to 599 ms, while 200 MB raises it to 1195 ms. Under attack, unrelated requests may experience p90 latency up to 1.8 seconds and maximum latency up to 2.5 seconds.

Useful 0 Not Useful 0

Mitigation Strategies

Upgrade vLLM to version 0.29.0 or later. Implement HTTP request size limits on your API endpoints to prevent excessively large payloads. Add input validation to enforce maximum length restrictions on the cache_salt parameter.

Useful 0 Not Useful 0