What is CVE-2026-100654?

CVE-2026-100654 is vLLM before 0.29.0 Denial of Service via out-of-range stop_token_ids. The supplied enrichment data classifies it as Denial of Service mapped to CWE-129 CWE-129.

Which versions are affected?

Affected product and version evidence is listed in the product table above. Current public advisories should be checked for exact version scope before patch targeting is finalized.

Is there active exploitation?

Active exploitation is not confirmed from current sources for CVE-2026-100654. If exploitation is not confirmed from current sources, teams should document that status and continue monitoring KEV, vendor, and telemetry updates.

No. CVE-2026-100654 is not listed in CISA KEV in the supplied enrichment data. KEV status should be monitored because it can change after initial disclosure or after exploitation becomes publicly confirmed.

Is there a public exploit?

1 public PoC repositories reported; 0 marked weaponized in current dataset. Public exploit or PoC availability should raise triage priority, especially for exposed or business-critical systems.

How can defenders detect it?

Start with web/proxy/WAF logs, application logs, network IDS/IPS where available. Look for vulnerability-class behavior, product errors, suspicious authentication events, exploit probes, and abnormal process activity based only on evidence that applies to the affected product.

How should teams prioritize it?

Prioritize CVE-2026-100654 using CVSS 6.5, EPSS 0%, KEV status, public PoC status, and asset exposure. Internet-facing, privileged, or business-critical deployments should be handled first.

Are workarounds available?

The supplied mitigation guidance says to apply vendor patches and restrict external exposure until fixed. Fixed-version evidence is listed in the mitigation section above.

Which threat actors are linked publicly?

No public threat actor linkage is confirmed from current sources. Do not attribute exploitation to a threat actor unless the source evidence explicitly supports that linkage.

What should be done in the next 24 hours?

Identify exposed assets running affected vendor/product/version combinations. Prioritize based on EPSS, PoC availability, and external exposure. Search available logs for exploit probes, errors, authentication anomalies, or suspicious child processes matching the vulnerability class. Open remediation tickets with affected assets, owners, evidence status, and detection coverage so decisions are traceable.

mediumDenial of Servicepublic exploit

CVE-2026-100654

CVE-2026-100654 — vLLM before 0.29.0 Denial of Service via out-of-range stop_token_ids

vLLM before 0.29.0 accepts user-controlled stop_token_ids on the OpenAI-compatible POST /v1/completions and POST /v1/chat/completions endpoints but validates only that the values are integers, not that each token id is within the model vocabulary/logits range. When min_tokens > 0, the stop token ids are used as logits indices to suppress stop tokens, so an out-of-range id reaches a CUDA indexing operation (index_put_) and triggers a device-side assertion. An authenticated API user can send a single malformed completion request that returns 500 Internal Server Error and puts EngineCore into a fatal state, causing subsequent requests to fail until the service is restarted (denial of service).

Published Sep 26, 2026Updated Sep 26, 2026Sources: NVD, vllm-project (CNA), GitHub, Red Hat (VEX)

Triage

Is it exploited, how likely is exploitation, what does it touch, and how severe do the scoring sources call it.

Exploitation

Exploit code

public exploit, none observed

EPSS

Not scored

FIRST has not published a score

Affects

vllm-project

2 products listed

CVSS base

6.5

medium · 2 sources

Remediation

The vendor's own words where we have them.

Upgrade vllm to a fixed release (0.29.0). Apply vendor patches per advisory and restrict external exposure of the affected component until patched.

Red Hat VEX · CVE-2026-100654

fixed0affected27not affected0

none_available: Fix deferred

First 24 hours

Ordered from the record's own fields — exposure first, because you cannot patch what you have not found.

Identify exposed assets running affected vendor/product/version combinations.
Prioritize based on EPSS, PoC availability, and external exposure.
Search available logs for exploit probes, errors, authentication anomalies, or suspicious child processes matching the vulnerability class.

Affected scope

Vendor, product and version as the advisories word them.

Vendor	Product	Versions	Status
vllm-project	vllm	0 to < 0.29.0	Vulnerable
vllm-project	vllm	0.29.0	Patched

Attack characteristics

The CVSS vector, decoded. It describes the attack, not your exposure to it.

Availability

High

Confidentiality

None

Integrity

None

Scope

Unchanged

Attack Complexity

Low

Attack Vector

Network

Privileges Req

Low

User Interaction

None

Every base score collected

Sources score independently and disagree; each row says who scored it and under which version.

Score	Version	Severity	Expl.	Impact	Source
6.5	CVSS 3.1	medium	2.8	3.6	NVD, vllm-project (CNA), Red Hat (VEX)
7.1	CVSS 4.0	high	—	—	vllm-project (CNA)

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

Weakness & attack patterns

CWE-129

Inferred from the weakness class — not observed against this CVE.

CAPEC-125 · FloodingCAPEC-130 · Excessive Allocation

Public exploit

Capability, not use: code existing is a different claim from anyone running it.

Repositories

Detection

Read off the CVSS vector and the weakness class. Starting points, not rules we have tested.

Search application, proxy, and WAF logs for requests touching /v1/completions, /v1/chat/completions, /logits.
Prioritize edge telemetry for network-reachable vllm.
Monitor for scanner or exploit-pattern traffic after 1 public PoC repositories were reported.

web/proxy/WAF logsapplication logsnetwork IDS/IPS

Timeline

What happened to this CVE, newest first — with the readings a source repeats on a schedule counted underneath rather than listed.

2026
Added · CVSS 3.1 6.5 (AV:N)
Sep 26, 2026 · NVD
Initial · CVE published
Sep 26, 2026 · NVD

References

6 on the record

github.com/vllm-project/vllm/security/advisories/GHSA-v5gm-qgmv-gc6c
Exploit, Third Party Advisory
www.vulncheck.com/advisories/vllm-before-0.29.0-denial-of-service-via-out-of-range-stop-token-ids
Vendor Advisory
access.redhat.com/security/cve/CVE-2026-100654
Vendor Advisory
bugzilla.redhat.com/show_bug.cgi?id=2541916
Vendor Advisory
www.cve.org/CVERecord?id=CVE-2026-100654
Vendor Advisory
nvd.nist.gov/vuln/detail/CVE-2026-100654
NVD

Elsewhere on this site

vllm-projectevery CVE for this vendor
Denial of Servicesame class
CWE-129other pages naming this weakness

Not in any source we poll

Listed rather than left blank: an empty field and an unmeasured one look identical on screen, and only one is a reason to look elsewhere.

No confirmed IOCs, IP addresses, domains, file hashes, or malware artifacts supplied.
No organization-specific asset inventory, compensating-control status, or patch deployment evidence supplied.
No exploit packet captures, log samples, or incident case IDs supplied.

Answered from this record1

What should defenders know first?

CVE-2026-100654 is vLLM before 0.29.0 Denial of Service via out-of-range stop_token_ids, a medium vulnerability affecting vllm from vllm-project. The current evidence does not list it in CISA KEV, and the exploit status is: Active exploitation is not confirmed from current sources for CVE-2026-100654. Public exploit evidence is: 1 public PoC repositories reported; 0 marked weaponized in current dataset. The affected-version evidence is listed in the key facts and affected products tables. Defenders should first verify whether exposed or business-critical assets run those versions, then apply vendor patches or mitigations, restrict reachable attack surface, and preserve logs for detection review. CVSS 6.5 describes technical severity, while EPSS 0% helps estimate near-term exploit likelihood; neither replaces asset context. Unknown fields should remain explicit in tickets, and threat actor, IOC, victimology, or payload claims should not be added unless a cited source supports them. Monitor CISA KEV, vendor advisories, NVD changes, public PoC repositories, and internal telemetry for update triggers.