vLLM: Harmony tool continuations drop `cache_salt` β€” restoring a cross-tenant prefix-cache membership oracle (vllm-project vllm)

CVE-2026-105752
LOW 3.1

At a glance

Source
NVD
EPSS
EPSS 0.22 % Β· percentile 12 ()
State
PUBLISHED
Published
Updated
Assigner
GitHub_M
NVD status
Analyzed

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.

Weaknesses (CWE)

  • CWE-200 β€” Exposure of Sensitive Information to an Unauthorized Actor
  • CWE-524 β€” Use of Cache Containing Sensitive Information

References

Scores

VersionScoreVectorScored by
CVSS 3.1LOW 3.1CVSS:3.1/AV:N/AC:H/PR:L/UI:N/S:U/C:N/I:L/A:NGitHub_M
CVSS 3.1LOW 3.1CVSS:3.1/AV:N/AC:H/PR:L/UI:N/S:U/C:N/I:L/A:NNVD Β· security-advisories@github.com (Secondary)

Advisories

European references

Affected configurations (NVD CPE)

Affected products

ProductVersions
vllm-project Β· vllmaffected < 0.30.0

Sources

SourceLicence
CVE List (CVEProject cvelistV5)CVE Program Terms of Use
EPSS (FIRST.org)FIRST β€” attribution requested
ENISA EUVDENISA β€” reproduction authorised, source acknowledged
NVD (NIST) via fkie-cad/nvd-json-data-feedsPublic domain β€” this product uses data from the NVD API but is not endorsed or certified by the NVD