
Cloud Vulnerability DB
A community-led vulnerabilities database
CVE-2026-111392 is a KV cache poisoning vulnerability in SGLang, a fast serving framework for large language models, affecting versions 0.5.13 through 0.5.21. The flaw allows remote unauthenticated attackers to poison the multimodal prefix-cache by supplying arbitrary mm_hashes values via the native /generate endpoint, enabling cross-tenant image content disclosure or response corruption. It was disclosed on October 16, 2026, by VulnCheck and assigned CVSS v3.1 score of 4.8 (Medium) and CVSS v4.0 score of 6.3 (Medium) (SGLang Issue #44641, VulnCheck Advisory).
The root cause is CWE-345 (Insufficient Verification of Data Authenticity): in tokenizer_manager.py, the TokenizerManager applies caller-supplied mm_hashes values unconditionally to each MultimodalDataItem via item.set_hash(int(hex_hash, 16)), bypassing the internal content-based hash_feature() recompute. Because MultimodalDataItem.hash determines the item's pad_value in schedule_batch.py, and the image region of the token sequence is that pad_value repeated, the declared hash becomes the prefix-cache identity for the image segment. A request sending image B while declaring image A's hash will match A's cached KV entries in the shared radix cache, causing the model to respond based on A's content regardless of the actual image bytes submitted — with no error or warning in the response (SGLang Issue #44641, tokenizer_manager.py, io_struct.py). The vulnerability is limited to the native /generate endpoint; OpenAI-compatible endpoints do not expose the mm_hashes field.
An unauthenticated remote attacker can disclose image content from other users' cached requests (confidentiality impact) or corrupt the model responses delivered to victim users by injecting stale or attacker-chosen cached KV entries (integrity impact). Because the radix cache is shared across all tenants, the attack is cross-tenant: a second client on a different connection can receive another client's image content in its response, as demonstrated in the proof-of-concept where cached_tokens jumped from 15 to 193, indicating the 162-token image region was served from a poisoned cache entry. Availability is not directly impacted (SGLang Issue #44641).
A proof-of-concept reproducer is publicly available on GitHub, consisting of a Python script that sends crafted HTTP POST requests to the /generate endpoint with manipulated mm_hashes values to trigger cross-tenant image disclosure (SGLang Issue #44641). The vulnerability requires no authentication and no user interaction, but exploitation requires high attack complexity — specifically, the attacker must know or predict a hash value previously cached by another request. The EPSS score is 0.0 and there is no evidence of in-the-wild exploitation or CISA KEV catalog listing as of the disclosure date. No threat actor attribution has been reported.
/generate endpoint (e.g., http://<host>:8200/generate). Confirm it is a multimodal deployment by checking for vision model indicators.mm_hashes value (e.g., deadbeef00000001) to populate the shared radix cache with that image's KV entries.image_data), the same mm_hashes value used in step 2, and a byte-identical prompt text (including the vision placeholder for the target model, e.g., <|vision_start|><|image_pad|><|vision_end|> for Qwen2.5-VL)./generate. The server applies the declared hash unconditionally, matches the cached KV from the original image, and returns a completion based on the cached image's content rather than the submitted image.text reflects the original (victim's) image content, and meta_info.cached_tokens shows a large cache hit (e.g., 193 vs. baseline 15), confirming successful cache poisoning and cross-tenant image disclosure (SGLang Issue #44641)./generate endpoint with a mm_hashes field containing hex strings that do not correspond to the submitted image content; requests from multiple source IPs using identical mm_hashes values.cached_tokens values (e.g., 162+ tokens) for image regions in requests where cache hits are not expected; absence of any error or warning in responses despite mismatched image/hash pairs (in unpatched versions).WARNING Ignoring caller-provided mm_hashes: --trust-caller-mm-hashes is not set indicate that callers are attempting to supply mm_hashes values, which may warrant investigation (SGLang PR #44653).The fix is implemented in SGLang PR #44653, which gates caller-declared mm_hashes behind a new opt-in server flag --trust-caller-mm-hashes. By default (flag unset), caller-supplied mm_hashes values are ignored and the content-based hash is recomputed internally, preventing cache poisoning. Operators who require the external KV router use case can explicitly set --trust-caller-mm-hashes at server launch, accepting the associated risk. Until a patched release is available, operators should restrict access to the native /generate endpoint to trusted clients only, or deploy network-level controls to prevent untrusted callers from supplying the mm_hashes field (SGLang PR #44653, SGLang Issue #44641).
The vulnerability was reported by researcher yyymk via a detailed GitHub issue with a complete reproducer. SGLang contributor Kylinny responded promptly on the same day, acknowledging the cross-tenant disclosure risk and opening PR #44653 with a fix within hours of the report. The fix design mirrors a precedent set by PR #34411 on the Kimi embedding path, which had already rejected caller mm_hashes from redirecting cache identity. No broader media coverage or significant social media discussion has been identified beyond the GitHub thread (SGLang Issue #44641, SGLang PR #44653).
Source: This report was generated using AI
Free Vulnerability Assessment
Evaluate your cloud security practices across 9 security domains to benchmark your risk level and identify gaps in your defenses.
Get a personalized demo
"Best User Experience I have ever seen, provides full visibility to cloud workloads."
"Wiz provides a single pane of glass to see what is going on in our cloud environments."
"We know that if Wiz identifies something as critical, it actually is."