CVE-2026-100652: vLLM: invalid stop_token_ids crash EngineCore (DoS)

HIGH
Published September 26, 2026
CISO Take

CVE-2026-100652 is a missing input validation flaw in vLLM 0.22.0 through 0.23.0 where the Rust HTTP and gRPC frontends do not check stop_token_ids against vocabulary bounds, so a single crafted request with min_tokens above zero and an out-of-vocabulary token ID triggers a CUDA indexing failure that leaves EngineCore fatally broken until the service is restarted. The attack is unauthenticated, network-reachable and low complexity (CVSS 7.5, availability only), so any internet-exposed or multi-tenant vLLM endpoint can be knocked offline by one request. Near-term mass exploitation looks limited: EPSS is 0.36% (72nd percentile), the CVE is not in CISA KEV, there is no public exploit or Nuclei template, and CISA SSVC says TRACK. However, vLLM has 95 downstream dependents and 134 other CVEs on record, so the blast radius for shared inference platforms is meaningful. Upgrade past 0.23.0 to the fixed release named in GHSA-qff2-492f-9fm4 (the affected-version range is documented but no patched version was listed in our data, so confirm it in the advisory). Until then, keep vLLM endpoints behind an authenticated gateway, validate stop_token_ids and min_tokens at the proxy, and set up automatic restart with alerting on EngineCore failures.

Sources:NVDEPSSGitHub Advisoryvulncheck.comATLASCISA SSVC

What is the risk?

Rated high, driven by trivial network exploitation with no privileges or user interaction, and an availability-only impact (C:N/I:N/A:H). Exploitation likelihood is currently low to moderate (EPSS 0.00361, 72nd percentile; no KEV listing, no public exploit, no scanner template; SSVC TRACK). Risk rises sharply for exposed or multi-tenant inference endpoints and for production services without automatic recovery, because the fatal EngineCore state needs a restart rather than self-healing. There is no confidentiality or integrity impact. Only deployments running the Rust HTTP or gRPC frontends on 0.22.0-0.23.0 are affected.

How does the attack unfold?

Initial Access
Attacker reaches an exposed or poorly authenticated vLLM HTTP/gRPC inference endpoint over the network.
AML.T0040
Exploitation
Attacker submits a request with min_tokens > 0 and out-of-vocabulary stop_token_ids that bypass frontend validation.
AML.T0049
Impact
MinTokensLogitsProcessor triggers a CUDA indexing failure, EngineCore enters a fatal state, and inference stops until restart.
AML.T0029

What systems are affected?

PackageEcosystemVulnerable RangePatched
vLLMpip—No patch
96.9K93 dependentsPushed 3d ago26% patched~37d to patchFull package profile →

Do you use vLLM? You're affected.

How severe is it?

CVSS 3.1
7.5 / 10
EPSS
0.4%
chance of exploitation in 30 days
Higher than 37% of all CVEs
Exploitation Status
No known exploitation
Sophistication
Trivial

What is the attack surface?

AVACPRUISCIA
AVNetwork
ACLow
PRNone
UINone
SUnchanged
CNone
INone
AHigh

What should I do?

1 step
  1. 1) Upgrade vLLM beyond 0.23.0 to the patched release listed in GHSA-qff2-492f-9fm4 and the VulnCheck advisory, and verify the fix version there. 2) Until patched, restrict access to the Rust HTTP/gRPC frontends to trusted networks and require authentication through an API gateway or reverse proxy. 3) Add gateway-level validation that rejects requests whose stop_token_ids fall outside the model's vocabulary size, or that combine min_tokens > 0 with untrusted stop_token_ids. 4) Apply per-client rate limiting. 5) Run vLLM under a supervisor (Kubernetes liveness/readiness probes, systemd) so a fatal EngineCore state triggers automatic restart. 6) Detect by alerting on EngineCore fatal errors, CUDA indexing or device-side assert messages in logs, repeated pod restarts, and inference health check failures; review request logs for out-of-range stop_token_ids.

What does CISA's SSVC say?

DecisionTrack
Exploitationnone
AutomatableNo
Technical Impactpartial

Source: CISA Vulnrichment (SSVC v2.0). Decision based on the CISA Coordinator decision tree.

How is it classified?

Which compliance frameworks are affected?

This CVE is relevant to:

EU AI Act
Article 15-Accuracy, robustness and cybersecurity
ISO 42001
A.6.2.6-AI system operation and monitoring
NIST AI RMF
MEASURE 2.7-AI system security and resilience are evaluated
OWASP LLM Top 10
LLM10:2025-Unbounded Consumption

Frequently Asked Questions

What is CVE-2026-100652?

CVE-2026-100652 is a missing input validation flaw in vLLM 0.22.0 through 0.23.0 where the Rust HTTP and gRPC frontends do not check stop_token_ids against vocabulary bounds, so a single crafted request with min_tokens above zero and an out-of-vocabulary token ID triggers a CUDA indexing failure that leaves EngineCore fatally broken until the service is restarted. The attack is unauthenticated, network-reachable and low complexity (CVSS 7.5, availability only), so any internet-exposed or multi-tenant vLLM endpoint can be knocked offline by one request. Near-term mass exploitation looks limited: EPSS is 0.36% (72nd percentile), the CVE is not in CISA KEV, there is no public exploit or Nuclei template, and CISA SSVC says TRACK. However, vLLM has 95 downstream dependents and 134 other CVEs on record, so the blast radius for shared inference platforms is meaningful. Upgrade past 0.23.0 to the fixed release named in GHSA-qff2-492f-9fm4 (the affected-version range is documented but no patched version was listed in our data, so confirm it in the advisory). Until then, keep vLLM endpoints behind an authenticated gateway, validate stop_token_ids and min_tokens at the proxy, and set up automatic restart with alerting on EngineCore failures.

Is CVE-2026-100652 actively exploited?

No confirmed active exploitation of CVE-2026-100652 has been reported, but organizations should still patch proactively.

How to fix CVE-2026-100652?

1) Upgrade vLLM beyond 0.23.0 to the patched release listed in GHSA-qff2-492f-9fm4 and the VulnCheck advisory, and verify the fix version there. 2) Until patched, restrict access to the Rust HTTP/gRPC frontends to trusted networks and require authentication through an API gateway or reverse proxy. 3) Add gateway-level validation that rejects requests whose stop_token_ids fall outside the model's vocabulary size, or that combine min_tokens > 0 with untrusted stop_token_ids. 4) Apply per-client rate limiting. 5) Run vLLM under a supervisor (Kubernetes liveness/readiness probes, systemd) so a fatal EngineCore state triggers automatic restart. 6) Detect by alerting on EngineCore fatal errors, CUDA indexing or device-side assert messages in logs, repeated pod restarts, and inference health check failures; review request logs for out-of-range stop_token_ids.

What systems are affected by CVE-2026-100652?

This vulnerability affects the following AI/ML architecture patterns: model serving, LLM inference endpoints, RAG pipelines, agent frameworks, multi-tenant GPU inference clusters.

What is the CVSS score for CVE-2026-100652?

CVE-2026-100652 has a CVSS v3.1 base score of 7.5 (HIGH). The EPSS exploitation probability is 0.42%.

What is the AI security impact?

Affected AI Architectures

model servingLLM inference endpointsRAG pipelinesagent frameworksmulti-tenant GPU inference clusters

MITRE ATLAS Techniques

AML.T0029Denial of AI Service
AML.T0040AI Model Inference API Access
AML.T0049Exploit Public-Facing Application

Compliance Controls Affected

EU AI Act: Article 15
ISO 42001: A.6.2.6
NIST AI RMF: MEASURE 2.7
OWASP LLM Top 10: LLM10:2025

What are the technical details?

Original Advisory

vLLM versions 0.22.0 through 0.23.0 fail to validate stop_token_ids against vocabulary bounds in Rust HTTP and gRPC frontends, allowing out-of-vocabulary token IDs to reach MinTokensLogitsProcessor. Attackers can submit requests with min_tokens greater than zero and out-of-vocabulary stop_token_ids to trigger CUDA tensor indexing failures that leave EngineCore in a fatal state requiring service restart.

Exploitation Scenario

An attacker finds an exposed vLLM completion endpoint (or an internal one reached via a compromised workload or an unauthenticated API gateway). They send a single request with min_tokens set to 1 or more and stop_token_ids containing a value far above the model's vocabulary size. The frontend passes it through unchecked, MinTokensLogitsProcessor performs an out-of-bounds CUDA index, and EngineCore enters a fatal state. The inference service stops responding to all users. The attacker resends the request each time the service restarts, sustaining the outage against a production AI assistant or customer-facing LLM API at negligible cost.

Weaknesses (CWE)

CWE-20 — Improper Input Validation: The product receives input or data, but it does not validate or incorrectly validates that the input has the properties that are required to process the data safely and correctly.

  • [Architecture and Design] Consider using language-theoretic security (LangSec) techniques that characterize inputs using a formal language and build "recognizers" for that language. This effectively requires parsing to be a distinct layer that effectively enforces a boundary between raw input and internal data representations, instead of allowing parser code to be scattered throughout the program, where it could be subject to errors or inconsistencies that create weaknesses. [REF-1109] [REF-1110] [REF-1111]
  • [Architecture and Design] Use an input validation framework such as Struts or the OWASP ESAPI Validation API. Note that using a framework does not automatically address all input validation problems; be mindful of weaknesses that could arise from misusing the framework itself (CWE-1173).

Source: MITRE CWE corpus.

CVSS Vector

CVSS:3.1/AV:N/AC:L/PR:N/UI:N/S:U/C:N/I:N/A:H

Timeline

Published
September 26, 2026
Last Modified
October 6, 2026
First Seen
September 26, 2026

Related Vulnerabilities