Vulnerability DatabaseCVE-2026-112266

CVE-2026-112266: 
SGLang vulnerability analysis and mitigation

Overview

CVE-2026-112266 is a denial-of-service vulnerability in SGLang (sgl-project/sglang) through version 0.5.22, caused by improper partial string comparison (CWE-187) in the scheduler's abort logic. An unauthenticated remote attacker can supply a short, client-controlled request ID (rid) to abort other users' in-flight generation requests via prefix-matching logic. The vulnerability was first privately reported on August 2, 2026 as GHSA-c7p3-8wqm-7cq3, publicly disclosed via GitHub issue #41474 on September 27, 2026, and formally published as a CVE on October 19, 2026. It carries a CVSS v3.1 base score of 5.9 (Medium) and a CVSS v4.0 base score of 8.2 (High) (GitHub Advisory, VulnCheck Advisory).

Technical details

The root cause is a logic flaw (CWE-187: Partial String Comparison) in Scheduler.abort_request() within python/sglang/srt/managers/scheduler.py. The scheduler matches abort targets using req.rid.startswith(recv_req.rid) instead of exact equality, meaning any abort request with a short rid will cancel all in-flight requests whose rid begins with that prefix. When a streaming /generate request is submitted with a client-supplied rid and the client disconnects, TokenizerManager.create_abort_task() schedules a background task that sleeps two seconds and then calls abort_request() with the attacker-controlled rid. Because TokenizerManager.abort_request() only checks for exact rid presence in rid_to_state before dispatching — and the long attacker request remains active during the 2-second delay — the prefix-match abort propagates to all queued, waiting, grammar, chunked, and running requests whose rids share the prefix. Default UUID rids (32 hex characters) are fully vulnerable to single-character hex prefix attackers (0–f) (GitHub Issue #45002, GitHub Issue #41474).

Impact

Successful exploitation causes a complete denial of service for all concurrent users of the SGLang inference server. An unauthenticated attacker can repeatedly abort other clients' generation requests — including those using default UUID-based rids — by cycling through 16 single hex-character prefixes (0–f), achieving 100% abort coverage of all default-rid requests. In testing, 40 out of 40 victim requests were aborted across five independent rounds using this technique, with aborts occurring approximately 2 seconds after attacker disconnect. The attack affects single-worker deployments and prefill-decode (PD) disaggregated deployments behind sglang-router, and there is no confidentiality or integrity impact — only availability (GitHub Issue #45002).

Exploitability

A proof-of-concept exploit is publicly available as a Python script in GitHub issue #41474, with detailed reproduction steps confirmed in issue #45002. The exploit requires no authentication, no special privileges, and no user interaction — only network access to the SGLang server (default port 30000). The EPSS score is 0.0042 (low probability of widespread exploitation), and there is no evidence of in-the-wild exploitation at this time. The CVE status is listed as "Deferred" and it is not currently in the CISA KEV catalog. NVD SSVC classifies exploitation status as "poc" (GitHub Issue #41474, GitHub Advisory).

Exploitation steps

  1. Reconnaissance: Identify an exposed SGLang server (default port 30000) running version ≤ 0.5.22 using network scanning tools or by checking the /health endpoint.
  2. Identify victim requests: Confirm other clients are submitting generation requests. Default UUID rids (32 hex chars) are vulnerable to single-character prefix attacks; sequential rids like job-10, job-11 are vulnerable to a job-1 prefix abort.
  3. Submit attacker streaming request: Send a POST request to /generate with a short rid (e.g., a single hex character like "a" or a shared prefix like "job-1") and "stream": true, with a long generation task to keep the request active on the server:
import json, urllib.request
payload = json.dumps({"text": "Count slowly...", "rid": "a", "stream": True, "sampling_params": {"max_new_tokens": 2048, "ignore_eos": True}}).encode()
req = urllib.request.Request("http://TARGET:30000/generate", data=payload, headers={"Content-Type": "application/json"}, method="POST")
response = urllib.request.urlopen(req)
  1. Disconnect the attacker stream: Close the HTTP connection after receiving the first token. The server retains the attacker's request in rid_to_state for ~2 seconds.
  2. Trigger prefix-match abort: After the 2-second delay, create_abort_task() fires abort_request("a"), which the scheduler applies via req.rid.startswith("a"), terminating all in-flight requests whose rid starts with "a".
  3. Repeat for full coverage: Cycle through all 16 single hex characters (0–f) to abort all requests using default UUID rids, achieving complete denial of service for all concurrent users (GitHub Issue #45002, GitHub Issue #41474).

Indicators of compromise

  • Network: Repeated short-lived POST requests to /generate on port 30000 with very short rid values (single characters or short strings); connections that close immediately after receiving the first streaming token.
  • Logs: SGLang server logs showing multiple requests finishing with finish_reason: "abort" or {"type": "abort", "message": "Aborted"} approximately 2 seconds after a streaming client disconnects; log entries like Abort queued request. req.rid=<victim_rid> in the scheduler debug log.
  • Behavioral: Legitimate long-running generation requests terminating prematurely (~2 seconds after an unrelated client disconnects); all concurrent requests aborting in rapid succession; server health checks remaining successful while generation requests consistently fail.

Mitigation and workarounds

Upgrade SGLang to a version beyond 0.5.22, which contains the fix replacing prefix-based startswith() comparisons with exact equality checks in Scheduler.abort_request(). As a workaround prior to patching, restrict network access to the SGLang API port (default 30000) to trusted clients only using firewall rules or an API gateway with authentication. Additionally, consider implementing server-side request ID generation (ignoring or namespacing client-supplied rids) and adding rate limiting on /generate requests to reduce the attack surface. The fix PRs referenced in the disclosure are #41481, #41502, and #43476 (GitHub Advisory, VulnCheck Advisory).

Community reactions

The vulnerability was first privately reported through GitHub Security Advisories on August 2, 2026 (GHSA-c7p3-8wqm-7cq3), but the private report was not processed, leading the reporter to disclose publicly via issue #45002 on October 18, 2026. A separate researcher independently discovered and reported the same root cause in issue #41474 on September 27, 2026. Community reaction focused on the practical severity of the bug in multi-tenant LLM serving environments, particularly given that default UUID rids are fully exploitable via 16 single-character prefix attacks. VulnCheck assigned and published the CVE (GitHub Issue #45002, VulnCheck Advisory).

Additional resources


Source: This report was generated using AI

Related SGLang vulnerabilities:

CVE ID

Severity

Score

Technologies

Component name

CISA KEV exploit

Has fix

Published date

CVE-2026-112266HIGH8.2
  • SGLang logoSGLang
  • cpe:2.3:a:lmsys:sglang
NoYesOct 19, 2026
CVE-2026-112270HIGH7.1
  • SGLang logoSGLang
  • cpe:2.3:a:lmsys:sglang
NoYesOct 19, 2026
CVE-2026-112269HIGH7.1
  • SGLang logoSGLang
  • cpe:2.3:a:lmsys:sglang
NoYesOct 19, 2026
CVE-2026-112267MEDIUM6.9
  • SGLang logoSGLang
  • cpe:2.3:a:lmsys:sglang
NoYesOct 19, 2026
CVE-2026-111392MEDIUM6.3
  • SGLang logoSGLang
  • cpe:2.3:a:lmsys:sglang
NoNoOct 16, 2026

Free Vulnerability Assessment

Benchmark your Cloud Security Posture

Evaluate your cloud security practices across 9 security domains to benchmark your risk level and identify gaps in your defenses.

Request assessment

Additional Wiz resources

Get a personalized demo

Ready to see Wiz in action?

"Best User Experience I have ever seen, provides full visibility to cloud workloads."
David EstlickCISO
"Wiz provides a single pane of glass to see what is going on in our cloud environments."
Adam FletcherChief Security Officer
"We know that if Wiz identifies something as critical, it actually is."
Greg PoniatowskiHead of Threat and Vulnerability Management