Medium severity · published Oct 16, 2026
CVE-2026-111392
SGLang through 0.5.21 KV Cache Poisoning via mm_hashes in /generate
4.8
CVSS
0.2%
EPSS
No
On KEV
Description
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.
CVSS vector
CVSS:3.1/AV:N/AC:H/PR:N/UI:N/S:U/C:L/I:L/A:N
Attack vector
Network
Complexity
High
Privileges
None
User interaction
None
Confidentiality
Low
Integrity
Low
Availability
None
Scope
Unchanged
Exploitation
Is it being used?
Exploit prediction (EPSS)0.2% · higher than 10% of CVEs
Exploited in the wildNo evidence yet
CISA KEVNot listed
Ransomware campaign useNot known
Affected products
Which versions
SG
sglang · sgl-project
0.5.13 through 0.5.21 · source: vendor
References
NVD
https://nvd.nist.gov/vuln/detail/CVE-2026-111392
CVE.org
https://www.cve.org/CVERecord?id=CVE-2026-111392
github.com
https://github.com/sgl-project/sglang/blob/v0.5.21/python/sglang/srt/managers/io_struct.py#L205-L215
github.com
https://github.com/sgl-project/sglang/blob/v0.5.21/python/sglang/srt/managers/tokenizer_manager.py#L1183-L1215
github.com
https://github.com/sgl-project/sglang/issues/44641
github.com
https://github.com/sgl-project/sglang/pull/44653
www.vulncheck.com
https://www.vulncheck.com/advisories/sglang-through-0.5.21-kv-cache-poisoning-via-mm-hashes-in-generate
More from sglang
8.6
CVE-2026-92972
SGLang through 0.5.19 Unauthenticated Route Poisoning via PUT endpoint
8.1
CVE-2026-111391
SGLang 0.5.10 through 0.5.22 Code Injection via Debug Dumper Filter
7.5
CVE-2026-112082
sgl-project SGLang Expert-Distribution Recording Endpoint expert_distribution.py Scheduler.expert_distribution_handle denial of service
7.5
CVE-2026-109229
SGLang through 0.5.21 Decompression Bomb DoS via NanoNemotronVLImageProcessor
7.5
CVE-2026-109228
SGLang through 0.5.21 Resource Budget Bypass via GLM-V Video Parameters