Attack ChainAIDerived
Hypothetical attack flow derived from CVE metadata – not a confirmed exploit chain
-
AccessReach exposed SGLang endpoint
-
DeliverySend concurrent oversized multipart uploads
-
ExploitFiles fully buffered in memory before validation
-
ExecutionExhaust worker memory
-
ImpactCrash HTTP/tokenizer worker (denial of service)
Vulnerability Assessment
| Weakness |
CWE
OWASP Top 10:2025
A06:2025 - Insecure Design
ASVS 5.0
V2 - Validation and Business Logic Resource allocation limits
MITRE ATT&CK
1 technique across 1 tactic: Impact The technique list with ATT&CK links and the kill-chain view is part of the Basic plan and up. |
| Exploitation Conditions | Exploitation requires that the target run SGLang (through 0.5.21) with its HTTP server exposing the /v1/audio/transcriptions endpoint reachable by the attacker; the endpoint accepts multipart file uploads without authentication (PR:N/UI:N) and buffers the entire uploaded file into memory before performing size validation, so sending concurrent oversized uploads triggers memory exhaustion. The practical limiting factors are network exposure (the SGLang API must be reachable rather than bound to localhost or firewalled) and the availability-only impact - successful attacks crash the worker and deny service but do not compromise data or allow code execution, and the service can be restarted. |
| Impact | All signals point to a genuine but impact-limited availability threat. The provided CVSS 4.0 vector (AV:N/AC:L/AT:N/PR:N/UI:N) confirms a network-reachable, low-complexity, unauthenticated attack with no user interaction, and the impact metrics (VC:N/VI:N/VA:H) correctly scope this as pure denial-of-service with no confidentiality or integrity loss - consistent with the CWE-770 (unbounded resource allocation) classification and the description of memory exhaustion crashing the HTTP/tokenizer worker. No EPSS score, KEV listing, or public POC was provided in the input, so there is no public exploit identified at time of analysis and no evidence of active exploitation; the finding originates from VulnCheck research with a vendor/advisory reference and an upstream GitHub issue, not from observed attacks. The 8.7 base score is driven entirely by the ease of triggering a service outage, not by any code-execution or data-exposure risk. Net assessment: a real, easily-reproduced DoS priority for anyone exposing SGLang's audio transcription endpoint to untrusted networks, but materially lower urgency than a comparably-scored RCE or data-breach flaw because the worst case is a recoverable crash. No conflicting signals were detected; the main gap is the absence of exploitation-probability (EPSS) and in-the-wild data. |
| Exploit ScenarioDerived | An attacker who can reach an exposed SGLang server (through 0.5.21) sends multiple concurrent large multipart file uploads to the unauthenticated /v1/audio/transcriptions endpoint. Because each oversized file is fully read into memory before any size validation, the server's memory is exhausted and the HTTP/tokenizer worker crashes, denying service to legitimate users. No public exploit code was identified in the provided data, but the attack requires only a standard HTTP client and no authentication. |
Get an e-mail when this CVE is exploited, patched or re-scored. Free account.
Alert meAnalysisAIv1 – Oct 12, 2026
Unauthenticated memory exhaustion in SGLang through 0.5.21 lets remote attackers crash the HTTP/tokenizer worker and deny service by sending concurrent oversized multipart uploads to the /v1/audio/transcriptions endpoint, which reads each uploaded file fully into memory before validating its size. The flaw is CWE-770 (allocation of resources without limits or throttling) and is scored 8.7 under CVSS 4.0, with the independent assessment confirming CVSS 3.1 AV:N/AC:L/PR:N/UI:N/C:N/I:N/A:H. No public exploit identified at time of analysis, and the impact is availability-only: no confidentiality or integrity loss, and the service can be restarted once the flood stops.
RemediationAI
No vendor-released patch version was identified in the available data at time of analysis, so track the upstream issue https://github.com/sgl-project/sglang/issues/43709 and the VulnCheck advisory https://www.vulncheck.com/advisories/sglang-through-0.5.21-memory-exhaustion-dos-via-v1-audio-transcriptions-upload for a fix and upgrade as soon as a patched release is published. Until then apply compensating controls: disable the /v1/audio/transcriptions endpoint entirely if transcription is not needed, since that removes the vulnerable code path completely; if it is needed, front SGLang with a reverse proxy (for example nginx client_max_body_size or an Envoy max request bytes limit) that rejects oversized bodies before they reach the worker, accepting that legitimate large audio files above the chosen limit will also be rejected; enforce per-client rate limiting and a cap on concurrent in-flight uploads to prevent the concurrency amplification that turns a single oversized file into memory exhaustion, with the trade-off of reduced throughput for legitimate batch transcription; restrict reachability by binding the API to internal interfaces and requiring authentication (API key or mTLS) at the proxy layer, since the endpoint itself performs no authentication; and run the service under memory cgroups or ulimits so a successful flood is contained to the worker rather than the host, recognizing that this causes the crash to occur sooner rather than preventing it. Monitor worker restarts and memory usage on the transcription endpoint as a detection signal, and restart the HTTP/tokenizer worker to restore service after an attack subsides.
Affected VersionsENISA
Affected Packages
Source: OSV (Open Source Vulnerabilities)
Affects your stack
Affects 1 package in euvd
Checking this CVE against your own stacks and servers is part of the Basic plan and up.
Technical ContextAI · show details
SGLang (sgl-project/sglang) is a high-throughput framework for serving large language models that exposes OpenAI-compatible REST endpoints, including /v1/audio/transcriptions for speech-to-text. In version 0.5.21 the handler implemented at python/sglang/srt/entrypoints/http_server.py lines 1847-1871 processes multipart file uploads by reading the entire request body into memory before applying any size check, so memory consumption scales directly with the attacker-supplied file size; concurrent uploads multiply these allocations across worker threads/processes until the HTTP/tokenizer worker is OOM-killed. The cited CWE-770 root cause is the absence of allocation limits or throttling: no maximum upload size, no streaming-to-disk/spooling, and no bound on concurrent in-flight requests on that endpoint. The CPE cpe:2.3:a:sgl-project:sglang:*:*:*:*:*:*:*:* with a wildcard version indicates the entire SGLang release line up to and including 0.5.21 is affected, and the CVE metadata carries only availability impact (VC:N/VI:N/VA:H), with no code execution or data exposure described.
Know when Sglang gets its next vulnerability
- Watch products and get an email when something new is published
- Check this CVE against your own stack and servers
- Exploit and PoC references, threat intelligence from 22 sources
No credit card · 7-day full trial
Severity by source
CVSS:4.0/AV:N/AC:L/AT:N/PR:N/UI:N/VC:N/VI:N/VA:H/SC:N/SI:N/SA:N
Unauthenticated, low-complexity network uploads to an open endpoint (AV:N/AC:L/PR:N/UI:N); only availability is affected via worker crash (A:H, C:N/I:N), no scope change.
Primary rating from Vendor (VulnCheck).
CVSS VectorVendor: VulnCheck
DescriptionCVE.org
SGLang through 0.5.21 contains an unbounded memory allocation vulnerability that allows unauthenticated remote attackers to exhaust memory by uploading oversized files to /v1/audio/transcriptions. Attackers can send concurrent large multipart uploads, which are fully read into memory before validation, to crash the HTTP/tokenizer worker and deny service.
Executive ViewAI
SGLang deployments up to and including version 0.5.21 are exposed to an unauthenticated remote denial-of-service condition in which concurrent oversized file uploads to the /v1/audio/transcriptions endpoint exhaust server memory and crash the HTTP/tokenizer worker. Because no authentication is required, any host that can reach the inference API can take it offline, disrupting any product, internal tool, or customer-facing service that depends on SGLang for model serving. Leadership should treat this as an availability and service-continuity risk rather than a data-loss event, but the absence of a vendor patch means exposure will persist until network and application-layer mitigations are in place.
No vendor-released patch identified at time of analysis. Interim mitigations include disabling or blocking the /v1/audio/transcriptions endpoint if not needed, enforcing strict request body size caps and per-client concurrency/rate limits at the reverse proxy or API gateway, restricting network access to trusted clients, applying container memory limits and automatic restart policies to contain worker crashes, and monitoring for memory exhaustion and service restarts.
Threat Intelligence Overview
How this decision was reached
The engine reads each signal below and applies a fixed list of rules from top to bottom. The first rule that matches sets the recommended action – like a statement, every line is shown.
What every action level means
SSVC AssessmentENISA
Priority Score
Attack Technique
Denial of Service attacks render applications or systems unavailable by overwhelming resources or triggering failure conditions.
Unauthenticated remote attackers can crash the SGLang inference server (versions up to and including 0.5.19) by calling
Remote code execution in SGLang (versions up to and including 0.5.15) allows unauthenticated attackers to run arbitrary
SGLang's multimodal generation module deserializes untrusted data with pickle.loads() over an unauthenticated ZMQ broker
SGLang's encoder parallel disaggregation system is vulnerable to unauthenticated RCE through pickle deserialization in t
Unauthenticated remote code execution in SGLang (the LLM/multimodal generation serving runtime) affecting version 5.10 a
Remote code execution in SGLang 0.5.9's /v1/rerank endpoint allows unauthenticated attackers to execute arbitrary code b
Remote code execution in SGLang (versions up to and including 0.5.15) allows attackers to run arbitrary code on the infe
Remote code execution in SGLang AI inference servers allows unauthenticated attackers to run arbitrary code through the
Remote code execution in SGLang (versions ≤ v0.5.15) allows attackers to achieve arbitrary code execution through the op
Unauthenticated remote code execution affects SGLang, an LLM/multimodal inference-serving framework, at version 5.10, wh
Unauthenticated remote code execution in SGLang versions through v0.5.20 arises from unsafe pickle deserialization in th
Unauthenticated remote code execution in SGLang 0.5.11 through 0.5.14 occurs when the multimodal generation runtime is l
EUVD-2026-97431
GHSA-4fxq-6r94-mchm
NVD
CVE.org
EPSS
Exploit-DB