Skip to main content
Vulnerability Database/CVE-2026-114049

CVE-2026-114049: vLLM Cross-Request Mask Leak Vulnerability

CVE-2026-114049 is a cross-request mask leak in vLLM that allows attackers to constrain other clients outputs through stale token IDs. 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-2026-114049 Overview

CVE-2026-114049 is a cross-request data leak in vLLM, an open-source inference and serving engine for large language models. Versions before 0.31.0 contain a flaw in InputBatch.condense() that leaves stale allowed_token_ids rows in recycled batch slots. Remote attackers can submit concurrent constrained generation requests, including through the unauthenticated /invocations endpoint, so a victim's unrestricted request inherits the attacker's token allowlist. The result is that one client can constrain another client's model outputs without authentication. The flaw is tracked under [CWE-440: Expected Behavior Violation].

Critical Impact

Unauthenticated remote attackers can manipulate the token distribution of concurrent users' LLM requests, degrading integrity of generated outputs in multi-tenant vLLM deployments.

Affected Products

  • vLLM versions prior to 0.31.0
  • Deployments exposing the /invocations HTTP endpoint
  • Multi-tenant inference services built on vulnerable vLLM releases

Discovery Timeline

  • 2026-10-21 - CVE-2026-114049 published to NVD
  • 2026-10-22 - Last updated in NVD database

Technical Details for CVE-2026-114049

Vulnerability Analysis

vLLM batches concurrent requests to maximize GPU throughput. Each request slot in InputBatch can carry per-request state, including an allowed_token_ids mask used for constrained decoding. When a request completes, InputBatch.condense() compacts the batch and recycles slots for new requests.

The condense routine fails to clear the previously populated allowed_token_ids rows for recycled slots. A new request that does not specify a token allowlist inherits the stale mask from the prior tenant. The sampler then applies that attacker-controlled allowlist to the victim's generation step, restricting which tokens the model may emit.

Because the /invocations endpoint is reachable without authentication in common deployment patterns, remote attackers can continuously submit constrained requests and poison recycled slots at scale.

Root Cause

The root cause is incomplete state reset during batch slot recycling. InputBatch.condense() moves active rows into compacted positions but does not zero or invalidate the trailing mask tensors associated with freed slots. The decoding pipeline treats the residual mask as valid, violating the expected per-request isolation boundary.

Attack Vector

An attacker submits concurrent constrained generation requests over the network, specifying a narrow allowed_token_ids list. As vLLM condenses the batch, the attacker's mask persists in slots later assigned to other clients. Victim requests that omit token constraints have their sampling restricted to the attacker's allowlist. The attack requires network reach to the vLLM API and concurrent traffic timing, but no credentials. Refer to the GitHub Security Advisory GHSA-6cxc-2vcg-w5qc for implementation specifics.

Detection Methods for CVE-2026-114049

Indicators of Compromise

  • Unexpected narrowing of model output vocabulary for legitimate client requests
  • High volumes of constrained generation requests containing unusual allowed_token_ids payloads to /invocations
  • Anomalous request bursts from a single source coinciding with user-reported output anomalies
  • Deployed vLLM package version lower than 0.31.0

Detection Strategies

  • Inventory running vLLM instances and compare installed versions against 0.31.0 using software composition analysis
  • Inspect reverse-proxy and application logs for unauthenticated POSTs to /invocations with constrained-decoding fields
  • Correlate spikes in constrained-generation traffic with downstream complaints of malformed or truncated model responses

Monitoring Recommendations

  • Enable structured request logging on vLLM frontends, capturing allowed_token_ids usage and client identity
  • Alert on bursts of concurrent constrained requests originating from the same IP or token
  • Monitor output token entropy per client; sustained drops may indicate a leaked mask applied to victim requests

How to Mitigate CVE-2026-114049

Immediate Actions Required

  • Upgrade vLLM to version 0.31.0 or later across all inference nodes
  • Place authentication and rate limiting in front of /invocations and other vLLM HTTP endpoints
  • Audit recent request logs for abusive constrained-generation patterns and notify affected tenants

Patch Information

The vLLM maintainers resolved the issue in release 0.31.0 by ensuring InputBatch.condense() fully clears allowed_token_ids rows for recycled slots. Review the GitHub Security Advisory GHSA-6cxc-2vcg-w5qc and the VulnCheck Advisory for vLLM Token IDs for upgrade guidance.

Workarounds

  • Disable constrained decoding features until upgrading if operational requirements allow
  • Restrict the /invocations endpoint to authenticated internal clients using a reverse proxy or API gateway
  • Isolate tenants into separate vLLM processes so recycled slots cannot cross trust boundaries
bash
# Upgrade vLLM to a patched release
pip install --upgrade "vllm>=0.31.0"

# Verify the installed version
python -c "import vllm; print(vllm.__version__)"

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.