Skip to main content
Vulnerability Database/CVE-2026-112266

CVE-2026-112266: SGLang DOS Vulnerability via Request ID

CVE-2026-112266 is a denial of service flaw in SGLang through version 0.5.22 that enables attackers to abort other users' requests. This post explains its technical details, affected versions, impact, and mitigation steps.

Published:

Disclaimer: This content was generated using AI. While we strive for accuracy, please verify critical information with official sources.

Sources: National Vulnerability Database, CISA KEV, Exploit-DB

CVE-2026-112266 Overview

CVE-2026-112266 is a denial of service vulnerability in SGLang through version 0.5.22. The flaw resides in the request identifier (rid) handling logic used by the streaming /generate endpoint. An unauthenticated API client can submit a streaming generation request using a short client-controlled rid, then disconnect. The server-side prefix-matching abort logic treats the disconnected short rid as a prefix match for other active rid values and terminates those unrelated requests. This cross-request abort behavior enables a single attacker to disrupt generation for all concurrent users of the inference server.

Critical Impact

An unauthenticated network attacker can abort other users' in-flight large language model (LLM) generation requests by submitting streaming requests with single-character hex rid values.

Affected Products

  • SGLang through version 0.5.22
  • SGLang streaming /generate API endpoint
  • Deployments exposing the SGLang tokenizer manager and scheduler to untrusted clients

Discovery Timeline

  • 2026-10-19 - CVE-2026-112266 published to the National Vulnerability Database (NVD)
  • 2026-10-19 - Last updated in NVD database

Technical Details for CVE-2026-112266

Vulnerability Analysis

SGLang is a serving framework for large language models that exposes an HTTP /generate endpoint supporting streaming responses. Each incoming request carries a request identifier called rid. The tokenizer manager and scheduler use the rid to correlate tokens, cache entries, and abort signals for in-flight generations.

The vulnerability stems from the abort routine using prefix matching rather than exact equality when a client disconnects. When a streaming client terminates its connection, the server iterates over active requests and aborts any entry whose rid begins with the disconnected client's rid. An attacker who submits a request with a single hex character (for example a) and disconnects causes the server to abort every active request whose rid starts with a. Across many short attacker-controlled rid prefixes, all concurrent requests can be terminated.

The weakness is categorized as [CWE-187] (Partial String Comparison). It does not affect confidentiality or integrity; the impact is availability loss for the inference service.

Root Cause

The abort logic in tokenizer_manager.py and scheduler.py compares each in-flight rid against the disconnecting client's rid using a prefix check. Because clients fully control the rid field in the request payload, nothing prevents an attacker from supplying a one-character value that acts as a prefix for arbitrary other identifiers.

Attack Vector

Exploitation requires only network access to the SGLang HTTP API. The attacker issues a streaming /generate request with a short rid such as a single hex character, then abruptly closes the TCP connection. The server executes its disconnect handler, which walks active generation requests and cancels those whose rid matches the attacker-supplied prefix. Repeating this request across the hex alphabet (0-f) covers the full space of auto-generated identifiers and reliably denies service to all other users. No authentication, elevated privileges, or user interaction is required.

For exploit mechanics and code locations, refer to the VulnCheck Security Advisory, the SGLang Scheduler source, and the Tokenizer Manager source.

Detection Methods for CVE-2026-112266

Indicators of Compromise

  • Streaming /generate requests containing unusually short rid values, especially single hex characters (0-9, a-f).
  • Rapid client disconnects immediately after submitting a streaming generation request.
  • Spikes in server-side abort events affecting multiple unrelated rid values within a narrow time window.
  • Clustered client errors or empty streaming responses reported across unrelated API tenants or users.

Detection Strategies

  • Log the raw rid value, source IP, and request lifetime for every /generate request and alert on rid lengths below a reasonable threshold (for example, fewer than 8 characters).
  • Correlate client disconnect events with downstream mass-abort events in the scheduler to detect the prefix-matching abuse pattern.
  • Baseline normal rid entropy and length produced by legitimate clients and flag deviations.

Monitoring Recommendations

  • Track the ratio of aborted-to-completed generations per source IP and alert when a single IP triggers aborts across many unrelated rid values.
  • Monitor API gateway metrics for short-lived streaming connections followed by elevated 5xx or stream termination events to downstream clients.
  • Forward SGLang application logs to a centralized analytics pipeline for cross-request correlation and historical review.

How to Mitigate CVE-2026-112266

Immediate Actions Required

  • Restrict network exposure of the SGLang /generate endpoint to trusted clients using firewall rules, a reverse proxy, or mutual TLS.
  • Enforce authentication and rate limiting at an API gateway in front of SGLang to constrain anonymous abuse.
  • Reject or rewrite client-supplied rid values at the proxy layer; generate server-side rid values with sufficient length and entropy.
  • Monitor upstream repositories for a fix beyond version 0.5.22 and plan an upgrade once a patched release is available.

Patch Information

As of publication, the vulnerability affects SGLang through version 0.5.22. Consult the SGLang project issues #41474 and #45002 along with the VulnCheck Security Advisory for upstream remediation status. Apply any released SGLang version that replaces prefix-based rid matching with exact equality checks.

Workarounds

  • Place SGLang behind a reverse proxy that overwrites the rid field with a server-generated UUID before the request reaches the tokenizer manager.
  • Apply strict input validation to reject rid values shorter than a defined minimum length (for example, 16 characters).
  • Deploy per-IP rate limits on the /generate endpoint to slow repeated short-request-and-disconnect patterns.
  • Isolate SGLang inference workloads per tenant so that a cross-request abort cannot span trust boundaries.

Disclaimer: This content was generated using AI. While we strive for accuracy, please verify critical information with official sources.

Sources: National Vulnerability Database, CISA KEV, Exploit-DB

Experience the Most Advanced Cybersecurity Platform

See how the world’s most intelligent, autonomous cybersecurity platform can protect your organization today and into the future.