GCD AI RADAR
incidents/AIR-2026-0239

vLLM tool continuations drop cache_salt to restore prefix-cache oracle

Attack on AI2 Disclosed vulnerabilityInfrastructure vulnerability

Tool continuations in vLLM fail to preserve cache_salt values, placing continuation prefixes in the global unsalted namespace. This allows authenticated tenants to determine prefix-cache membership of other users' histories.

disclosed CVENo confirmed exploitation, so this record is tracked as a disclosed vulnerability and not counted as an incident. It moves into the counts if CISA or the vendor confirms exploitation. Why

Details

type
Attack on AI
An AI model, agent or AI-powered product was the thing exploited, abused or leaked.
categories
  • Infrastructure vulnerability
    CVEs and advisories in AI frameworks, inference servers, MCP servers, AI developer tools and hosting stacks.
owasp llm top 10
mitre atlas
no technique for these categories
Both rows are derived from the type and categories (OWASP 2025 edition, ATLAS release 2026.09), not classified per record. How
severity
2 · Disclosed vulnerability
CVE or vendor advisory, patched or patchable, no confirmed exploitation.
targets
Framework / tooling
products
vLLM
incident date
5 Oct 2026
first reported
5 Oct 2026
origin
NVD
Vulnerability record from NVD / the GitHub Advisory Database. This product uses the NVD API but is not endorsed or certified by the NVD.
classifier confidence
0.95 · published at or above 0.80; see the methodology
tags
CVE-2026-105752 · side-channel · vllm
record
created 2026-10-18 06:33Z · updated 2026-10-18 06:33Z · json

Sources · 2

  1. 2026-10-05
    CVE-2026-105752: vLLM is an inference and serving engine for large language models.
    NVD · nvd.nist.gov/vuln/detail/CVE-2026-105752
  2. 2026-10-06
    CVE-2026-105752: vLLM: Harmony tool continuations drop `cache_salt` — restoring a cross-tenant prefix-cache membership oracle
    GitHub Advisory Database · github.com/advisories/GHSA-935w-9g4m-p28p

Titles and links belong to their publishers. The summary above is written by GCD AI RADAR in its own words; data is CC BY-SA 4.0.

Related incidents