{"id":"CVE-2026-111392","title":"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","summary":"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…","severity":"medium","cvss":4.8,"cvssVector":"CVSS:3.1/AV:N/AC:H/PR:N/UI:N/S:U/C:L/I:L/A:N","cwe":["CWE-345"],"vendor":"sgl-project","product":"sglang","affected":["sglang >= 0.5.13 <= 0.5.21"],"published":"2026-10-16","updated":"2026-10-16","sourceUpdated":"2026-10-16T20:17:20.370","source":"NVD","sourceUrl":"https://nvd.nist.gov/vuln/detail/CVE-2026-111392","references":[{"url":"https://github.com/sgl-project/sglang/blob/v0.5.21/python/sglang/srt/managers/io_struct.py#L205-L215","label":"disclosure@vulncheck.com"},{"url":"https://github.com/sgl-project/sglang/blob/v0.5.21/python/sglang/srt/managers/tokenizer_manager.py#L1183-L1215","label":"disclosure@vulncheck.com"},{"url":"https://github.com/sgl-project/sglang/issues/44641","label":"disclosure@vulncheck.com"},{"url":"https://github.com/sgl-project/sglang/pull/44653","label":"disclosure@vulncheck.com"},{"url":"https://www.vulncheck.com/advisories/sglang-through-0.5.21-kv-cache-poisoning-via-mm-hashes-in-generate","label":"disclosure@vulncheck.com"},{"url":"https://github.com/sgl-project/sglang/issues/44641","label":"134c704f-9b21-4f2e-91b3-4a467353bcc0"}],"tags":["nvd","cve.org","exploit-available"],"epss":0.00206,"epssPercentile":0.10495,"exploitAvailable":true,"ssvc":{"exploitation":"poc","automatable":"no","technicalImpact":"partial","timestamp":"2026-10-16T19:24:20.694655Z"},"ingestedAt":"2026-10-16T18:43:08.700Z","slug":"CVE-2026-111392","body":"## Overview\n\nSGLang 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.\n\n## Remediation\n\nRefer to the linked advisories for vendor-supplied fixes and affected version ranges.","depth":"twilight","depthScore":38,"depthScoreParts":{"impact":26.4,"likelihood":0,"exploitation":12,"ransomware":0},"changes":[{"seq":223341,"id":"CVE-2026-111392","ts":1792180099783,"field":"exploit_available","old":"false","new":"true"}]}