Attack ChainAIDerived
Hypothetical attack flow derived from CVE metadata – not a confirmed exploit chain
-
AccessReach SGLang /generate endpoint
-
DeliverySubmit streaming request with single hex rid
-
ExploitDisconnect to trigger abort
-
ExecutionPrefix-match terminates other users' requests
-
ImpactDenial of service for all clients
Vulnerability Assessment
| Weakness |
CWE
|
| Exploitation Conditions | Requires network reachability to SGLang's streaming /generate API (through version 0.5.22) and the ability to supply a short, client-controlled rid (e.g., a single hex character). Critically, impact is only realized when OTHER users' generation requests are concurrently in flight - the attacker aborts them via prefix-matching when disconnecting, so a single-tenant or idle server yields no victim impact (this is the basis for the vendor's AC:H/AT:P rating). No authentication is required (PR:N) and no user interaction is needed; the main limiting factors are the need for concurrent victim traffic and exposure of the /generate endpoint to the attacker (multi-tenant or internet-facing deployments are most at risk). |
| Impact | This is a genuine availability-focused issue but sits below the 'drop-everything' tier despite its 8.2 CVSS 4.0 score. The vendor vector (CVSS:4.0/AV:N/AC:H/AT:P/PR:N/UI:N/VC:N/VI:N/VA:H) is internally consistent with the description: network-reachable (AV:N), no authentication (PR:N), no user interaction (UI:N), and high availability impact (VA:H) with zero confidentiality/integrity impact - consistent with a pure DoS and the CWE-187 (partial/incorrect string comparison) root cause in the rid prefix-matching abort logic. The AC:H and AT:P metrics are the key risk dampeners: successful impact requires other users' requests to actually be in flight at the moment of the attack, so the attacker depends on conditions outside their control. No EPSS score, no CISA KEV listing, and no public POC are present in the provided data, so there is no public exploit identified at time of analysis - exploitation status is purely theoretical/vendor-reported (VulnCheck). Net: a real, low-effort-to-trigger availability bug worth patching on internet- or multi-tenant-exposed SGLang deployments, but not evidence of active, widespread exploitation. Note the absence of EPSS/KEV/POC is not itself a contradiction - it simply means no exploitation-probability signal is available to corroborate the high CVSS. |
| Exploit ScenarioDerived | An attacker with network access to a multi-tenant SGLang inference server submits streaming requests to the /generate endpoint using a single hex-character client-controlled rid, then disconnects. The disconnect triggers abort logic that prefix-matches the short rid against all active request IDs, terminating other legitimate users' in-flight generations and denying them service. No public exploit code is identified in the provided data, but the technique is fully described in the VulnCheck advisory. |
Get an e-mail when this CVE is exploited, patched or re-scored. Free account.
Alert meAnalysisAIv1 – Oct 19, 2026
Cross-request abort in SGLang through 0.5.22 lets unauthenticated remote API clients terminate other users' in-flight generation requests, producing a denial of service against multi-tenant or internet-facing inference servers. An attacker submits a streaming /generate request using a deliberately short, client-controlled rid (for example a single hex character) and then disconnects, at which point the server's prefix-matching abort logic cancels every concurrent request whose identifier begins with that character. The 8.2 CVSS 4.0 score is driven entirely by high availability impact with no confidentiality or integrity effect, and practical impact requires other tenants' requests to be in flight at the same moment; no public exploit code has been identified at time of analysis.
RemediationAI
No vendor-released patch was identified at time of analysis, so the primary action is to confirm a fixed SGLang release with the project (tracking issues 45002 and 41474 and the VulnCheck advisory at https://www.vulncheck.com/advisories/sglang-through-0.5.22-cross-request-abort-via-rid-prefix-matching) and upgrade as soon as one is published; do not assume a version beyond 0.5.22 is fixed without verification. Until then, the highest-value control is to remove attacker reachability to the streaming /generate endpoint: place it behind an authenticating reverse proxy or mTLS gateway and restrict it to known client networks, accepting the operational cost of credential distribution and the risk that any authenticated tenant can still trigger the abort. Where tenants are mutually untrusted, run separate SGLang instances (or separate scheduler/tokenizer processes) per tenant so that one client's rid cannot collide with another's in-flight identifiers - this costs GPU memory and duplication of model weights. Add a proxy or middleware layer that rejects /generate requests whose rid is shorter than a fixed minimum length or does not match an expected UUID/high-entropy format, which raises attacker effort but must be maintained in lockstep with any upstream rid scheme and can break legitimate clients that generate their own short IDs. Finally, alert on spikes of client disconnects correlated with very short rids and on bursts of cancelled requests, since that pattern is the observable signature of this attack.
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 is an LLM serving framework whose tokenizer manager and scheduler track in-flight generation work in shared, request-ID-keyed structures. When a client disconnects from a streaming /generate call, the abort path resolves which pending request to cancel by comparing the client-supplied rid against stored identifiers using a partial (prefix) comparison rather than exact equality - the weakness class assigned here is CWE-187, Partial Comparison. Because a one-character rid is a prefix of a large fraction of randomly generated identifiers, a single disconnect can fan out into cancellation of unrelated users' requests. The vulnerable logic lives in python/sglang/srt/managers/tokenizer_manager.py (the rid handling around lines 2240-2256 and the abort/cleanup path around lines 2425-2436) and in the request-abort handling in python/sglang/srt/managers/scheduler.py around lines 5458-5590 at tag v0.5.22. The assessed vector AV:N/AC:H/PR:N/UI:N/S:U/C:N/I:N/A:H reflects that the endpoint is network-reachable and needs no credentials or user interaction, but that the abort logic is only consequential when victim requests are concurrently in flight, which is why the vendor rates attack complexity high and attack requirements present (AT:P). Single-tenant or idle servers do not experience victim impact.
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:H/AT:P/PR:N/UI:N/VC:N/VI:N/VA:H/SC:N/SI:N/SA:N
Network-reachable unauthenticated endpoint (AV:N/PR:N) but impact depends on concurrent victim requests, so AC:H; pure availability loss means C:N/I:N/A:H.
Primary rating from Vendor (VulnCheck).
CVSS VectorVendor: VulnCheck
DescriptionCVE.org
SGLang through 0.5.22 contains a denial of service vulnerability that allows API clients to abort other users' in-flight generation requests by supplying a short client-controlled rid. Attackers can submit streaming /generate requests with single hex character rids and disconnect, triggering prefix-matching abort logic that terminates all other clients' requests.
Executive ViewAI
SGLang deployments through version 0.5.22 expose multi-tenant and internet-facing inference services to a denial-of-service condition that can terminate other customers' in-flight generation requests, directly threatening availability of AI inference workloads. Leadership should care because a single remote API client can disrupt concurrent tenant traffic, causing SLA breaches and erosion of trust in shared inference platforms, while no vendor patch is currently available.
No vendor patch; isolate tenants into separate SGLang instances, restrict API exposure to trusted networks, enforce server-side request ID generation at the gateway, and rate-limit or reject client-supplied short/prefix-colliding rid values.
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-2026-104307
GHSA-8v54-r9x8-p94v
NVD
CVE.org
EPSS
Vulnrichment
Exploit-DB