Vulnerability DatabaseCVE-2026-112267

CVE-2026-112267: 
SGLang vulnerability analysis and mitigation

Overview

CVE-2026-112267 is a Server-Side Request Forgery (SSRF) vulnerability in SGLang through version 0.5.22 when operating in prefill/decode disaggregation mode. Unauthenticated attackers can trigger outbound HTTP requests from the Decode worker by supplying attacker-controlled bootstrap_host and bootstrap_port fields in inference requests to the /generate endpoint. This allows internal network probing and can block the shared scheduler loop, causing severe latency increases for other inference requests. The vulnerability was disclosed on October 19, 2026, with a CVSS v3.1 score of 6.5 (Medium) and a CVSS v4.0 score of 6.9 (Medium) (Github Advisory, Feedly).

Technical details

The root cause is CWE-918 (Server-Side Request Forgery): in CommonKVManager.try_ensure_parallel_info() within python/sglang/srt/disaggregation/common/conn.py, the Decode worker constructs an HTTP URL directly from the client-supplied bootstrap_addr (derived from bootstrap_host and bootstrap_port request fields) without validating that the address belongs to a configured Prefill worker. The vulnerable code issues a synchronous requests.get(f"http://{bootstrap_addr}/route?...", timeout=5) call, which blocks the shared Decode request processing loop for the duration of the request. The attack surface is direct access to the Decode worker's HTTP API — the official Gateway overwrites these fields server-side via inject_bootstrap_into_value(), so the Gateway path is not exploitable. No authentication or privileges are required, and the attack is fully network-accessible (SGLang Issue #45008, conn.py source).

Impact

Successful exploitation enables two distinct impacts: internal network reconnaissance (SSRF) and denial of service. An attacker with direct network access to the Decode worker API can probe arbitrary internal hosts and ports by observing whether the worker's synchronous HTTP request succeeds or times out. Additionally, by directing the worker to a slow endpoint, the attacker can block the shared scheduler loop — measured experiments showed concurrent normal inference request latency increasing from ~0.246 seconds to ~20.354 seconds (an ~83x increase), though all requests ultimately returned HTTP 200 rather than failing outright. Confidentiality impact is limited (no direct data exfiltration), but the SSRF primitive could facilitate lateral movement into internal services reachable from the Decode worker (SGLang Issue #45008, Github Advisory).

Exploitability

A proof-of-concept Python reproduction script is publicly available in the SGLang GitHub issue tracker, containing a complete exploit harness with a canary HTTP server and concurrent request measurement capability (SGLang Issue #45008). The NVD SSVC assessment classifies exploitation as "poc" and the attack as "automatable" with no authentication required. The EPSS score is approximately 0.42% (37th percentile), indicating a relatively low but non-negligible probability of exploitation in the wild. There is no evidence of active in-the-wild exploitation as of the disclosure date, and the vulnerability is not listed in the CISA KEV catalog (Github Advisory, Feedly).

Exploitation steps

  1. Identify the attack surface: Confirm that the target SGLang deployment runs in prefill/decode disaggregation mode and that the Decode worker's HTTP API (typically on port 30001) is directly reachable from the network — the Gateway path is not exploitable as it overwrites bootstrap fields.
  2. Set up a canary server: Start a listener HTTP server on an attacker-controlled host/port (or loopback for testing) that logs incoming requests and optionally delays responses to amplify the DoS effect.
  3. Craft the malicious inference request: Construct a JSON POST body for the /v1/chat/completions or /generate endpoint that includes attacker-controlled fields alongside normal inference parameters:
{
  "model": "/models/target-model",
  "messages": [{"role": "user", "content": "reply OK"}],
  "max_tokens": 8,
  "temperature": 0,
  "bootstrap_host": "<attacker-controlled-host>",
  "bootstrap_port": <attacker-controlled-port>,
  "bootstrap_room": <unique-integer>
}
  1. Send the request directly to the Decode worker: POST the crafted payload to http://<decode-worker>:<port>/v1/chat/completions. The Decode worker's try_ensure_parallel_info() function will issue a synchronous GET http://<attacker-host>:<attacker-port>/route?prefill_dp_rank=-1&prefill_cp_rank=-1&target_tp_rank=-1&target_pp_rank=-1.
  2. Observe SSRF and/or DoS: Confirm the canary server receives the /route?... request (SSRF confirmed). For DoS, use a delayed canary response (e.g., 2 seconds per request) and send concurrent legitimate requests through the Gateway — their latency will be severely elevated while the scheduler loop is blocked.
  3. Probe internal network: Repeat with different bootstrap_host/bootstrap_port values targeting internal services; infer reachability based on response timing and error codes from the Decode worker (SGLang Issue #45008).

Indicators of compromise

  • Network: Unexpected outbound HTTP GET requests from the Decode worker to non-Prefill-worker endpoints, particularly to paths matching /route?prefill_dp_rank=-1&prefill_cp_rank=-1&target_tp_rank=-1&target_pp_rank=-1; outbound connections from the SGLang Decode worker process to unusual internal or external IP:port combinations.
  • Logs: SGLang Decode worker logs showing HTTP 500 responses after unusually long timeouts (14–30+ seconds) for inference requests; repeated connection attempts to the same non-configured bootstrap address; access logs showing direct POST requests to the Decode worker API (port 30001 or equivalent) rather than through the Gateway.
  • Application Behavior: Sudden and severe latency spikes (10x–100x normal) on concurrent inference requests processed through the Gateway without a corresponding increase in model load; inference requests returning HTTP 200 but with anomalously high elapsed times (~20 seconds vs. normal ~0.25 seconds).
  • Process: The SGLang Decode worker Python process making unexpected outbound TCP connections (observable via netstat, ss, or network monitoring tools) to hosts not in the configured Prefill worker set (SGLang Issue #45008).

Mitigation and workarounds

Upgrade SGLang to a version newer than 0.5.22, which contains the fix for this vulnerability (Github Advisory). As immediate workarounds: (1) implement network segmentation to ensure the Decode worker's HTTP API is not directly reachable by untrusted clients — route all inference traffic exclusively through the official Gateway, which overwrites bootstrap fields server-side; (2) apply firewall rules to restrict outbound HTTP connections from the Decode worker to only the configured Prefill worker addresses and ports; (3) consider disabling prefill/decode disaggregation mode entirely if it is not required for the deployment; (4) implement rate limiting and request validation on the /generate endpoint to reject requests containing bootstrap_host or bootstrap_port fields from untrusted sources (Feedly).

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