Vulnerability · Published Oct 5, 2026 · Assigned by GitHub_M

CVE-2026-105752vLLM: Harmony tool continuations drop `cache_salt` — restoring a cross-tenant prefix-cache membership oracle

Low 3.1

In plain English

A low-severity information exposure in vllm from vllm-project. There are no confirmed reports of exploitation so far. The chance it is exploited in the next 30 days is estimated at 0.2% (higher than 12% of all known vulnerabilities). Fixed versions are available - upgrade the affected package.

Severity (CVSS 3.1)3.1Low
Exploit odds (30 days)0.2%top 87.8% of all CVEs
Exploited in the wildNo reportsCISA SSVC: none
Attack toolingNoneno public tooling known

Official description

vLLM is an inference and serving engine for large language models. Prior to 0.30.0, Harmony tool continuations submitted through "POST /v1/responses" requests rebuild the next-turn engine input without preserving the cache_salt value, placing the continuation prefix in the global unsalted cache namespace even when the caller enabled salting. On deployments with prefix caching enabled, which is the default, an authenticated tenant who can reconstruct a victim's low-entropy post-tool history can submit the same continuation and use the cached_tokens_per_turn count to determine whether the prefix was previously processed, defeating the intended tenant isolation of salted prefix caching. This issue is fixed in version 0.30.0.

Affected packages & fixed versions

EcosystemPackageFixed inAdvisory
PyPIvllm0.30.0GHSA-935w-9g4m-p28p
PyPIvllm0.30.0PYSEC-2026-4308

Affected products (as reported)

VendorProductVersions
vllm-project vllm < 0.30.0

Weakness: CWE-200, CWE-524 · Vector: CVSS:3.1/AV:N/AC:H/PR:L/UI:N/S:U/C:N/I:L/A:N

References