Vulnerability in Sgl-Project Sglang
CVE-2026-111392
SGLang through 0.5.21 trusts caller-supplied mm_hashes values as the multimodal prefix-cache key, allowing remote attackers to poison cached KV entries across requests. Attackers can send a different image under a known or predicted hash via /generate to disclose another request's image content or corrupt victims' responses.
Published · last modified .
CVSS v3 metric
CVSS v3 base score 4.8 (Medium). Vector: CVSS:3.1/AV:N/AC:H/PR:N/UI:N/S:U/C:L/I:L/A:N.
EPSS exploit prediction
No EPSS score from FIRST.org yet. New CVEs usually get one within a few days; no score means no signal yet, not low risk. How EPSS works.
Affected products
Weakness classification (CWE)
References
- github.com/sgl-project/sglang/issues/44641 (issue-tracking)
- github.com/sgl-project/sglang/blob/v0.5.21/python/sglang/srt/managers/tokenizer… (technical-description)
- github.com/sgl-project/sglang/blob/v0.5.21/python/sglang/srt/managers/io_struct… (technical-description)
- github.com/sgl-project/sglang/pull/44653 (issue-tracking, patch)
- www.vulncheck.com/advisories/sglang-through-0.5.21-kv-cache-poisoning-via-mm-ha… (third-party-advisory)
Frequently asked questions
- What is CVE-2026-111392?
- CVE-2026-111392 is a medium-severity vulnerability in Sgl-Project Sglang, classified under Insufficient Verification of Data Authenticity. CVSS score: 4.8/10. Published 2026-10-16.
- How severe is CVE-2026-111392?
- Medium severity. CVSS v3 base score is 4.8 out of 10.