CVE-2026-54234

HIGHPre-NVD 7.57.5—
EchelonGraph scoreMEDIUM confidence

This high-severity CVE scores 7.5 under a secondary CVSS source (NVD's own analysis pending). EPSS exploit probability: 0.3%, top 74% of all CVEs by exploit prediction. GitHub Security Advisory data not yet ingested — confidence will rise once GHSA publishes (typical lag: hours to days for open-source ecosystem CVEs; never for infrastructure-only CVEs).

Triggered by: NVD CVSS baseline
Sources: epss, secondary
Trending — 3 sources updated this weekElevated
7.5
EchelonGraph verdictPlan a fixSerious severity, but no confirmed exploitation yet.
  • High severity, but no confirmed exploitation yet
CISA-KEV: Not listedEPSS: 0%CVSS: 7.5Exploit: NoneExposed: 0

No vendor fix yet — apply a workaround or compensating control (WAF / firewall / segmentation) and watch for a patch.

vLLM is a high-throughput and memory-efficient inference and serving engine for LLMs. Prior to 0.24.0, a frontend-legal multi-request speculative decoding workload can cause the rejection sampler to produce a recovered token equal to the model vocabulary size boundary value, which is then converted to negative one when the engine selects the next live token for a request and is written back into the drafter's input ids; that out-of-vocabulary value is later consumed by the model's embedding and attention path and crashes the engine worker with a GPU device-side assertion. The same triggering request sequence is reachable through the public gRPC Generate and Abort endpoints, so a remote client that can send generation requests can crash the shared engine worker, aborting concurrent requests and causing a service-wide denial of service for other clients of the deployment until the worker is restarted. This issue is fixed in version 0.24.0.

CVSS v3
7.5
EG Score
7.5(medium)
EPSS
26.3%
KEV
Not listed

Published

July 6, 2026

Last Modified

July 7, 2026

Advisory Details (3)

Auto-updated Jul 7, 2026
🔬 Proof of concept available. Patch available. Sources: github_commit, github_pr, github.
github Patch Available🟡 PoC Available

Remote DoS in vLLM via Invalid Recovered Token Reinjection · Advisory · vllm-project/vllm · GitHub

https://github.com/vllm-project/vllm/security/advisories/GHSA-8wr5-jm2h-8r4f
github_pr

[Security] Fix remote DoS via invalid recovered token reinjection

Fix merged in vllm-project/vllm PR #44744 on 2026-06-10 — awaiting tagged release

https://github.com/vllm-project/vllm/pull/44744
github_commit

commit 8a5cf1ccd65e (vllm-project/vllm)

Fix landed in vllm-project/vllm commit 8a5cf1ccd65e — awaiting tagged release

https://github.com/vllm-project/vllm/commit/8a5cf1ccd65e8ac7635c402c1ec0b08988bc26ca

Weakness Classification(2)

MITRE Common Weakness Enumeration — the root-cause categories this CVE belongs to.

Data Freshness Timeline

(refreshed 21× in last 7d / 21× in last 30d)

Each row is a source pipeline that fetched or updated this CVE on that date, with what changed. For example, "NVD update" means NVD published or revised its analysis for this CVE; "MITRE cvelistV5" means we ingested or refreshed it from the CNA feed. Most recent first.

  1. 2026-07-11 23:46 UTCEG score recompute
  2. 2026-07-11 12:34 UTCEG score recompute
  3. 2026-07-11 08:27 UTCEPSS rescore
  4. 2026-07-11 08:27 UTCEPSS rescore
  5. 2026-07-11 01:23 UTCEG score recompute
  6. 2026-07-10 14:12 UTCEG score recompute
  7. 2026-07-10 02:58 UTCEG score recompute
  8. 2026-07-09 19:10 UTCEPSS rescore
  9. 2026-07-09 19:10 UTCEPSS rescore
  10. 2026-07-09 15:47 UTCEG score recompute
  11. 2026-07-09 04:32 UTCEG score recompute
  12. 2026-07-08 17:21 UTCEG score recompute
  13. 2026-07-08 15:16 UTCEPSS rescore
  14. 2026-07-08 15:16 UTCEPSS rescore
  15. 2026-07-08 06:10 UTCEG score recompute
  16. 2026-07-07 19:00 UTCEG score recompute
  17. 2026-07-07 13:46 UTCEPSS rescore
  18. 2026-07-07 13:46 UTCEPSS rescore
  19. 2026-07-07 07:47 UTCEG score recompute
  20. 2026-07-06 20:31 UTCEG score recompute
  21. 2026-07-06 20:29 UTCMITRE cvelistV5first tracked

Related CVEs(same CWE)

Frequently asked(5)

What is CVE-2026-54234?
CVE-2026-54234 is a high vulnerability published on July 6, 2026. vLLM is a high-throughput and memory-efficient inference and serving engine for LLMs. Prior to 0.24.0, a frontend-legal multi-request speculative decoding workload can cause the rejection sampler to produce a recovered token equal to the model vocabulary size boundary value, which is then converted…
When was CVE-2026-54234 disclosed?
CVE-2026-54234 was first published in the National Vulnerability Database on July 6, 2026, with the most recent update on July 7, 2026. EchelonGraph re-ingests CVE updates from NVD on a 2-hour cycle, so this page reflects the latest published state.
Is CVE-2026-54234 actively exploited?
CVE-2026-54234 is not currently on CISA's Known Exploited Vulnerabilities catalog. FIRST EPSS estimates a 26.3% percentile likelihood of exploitation in the next 30 days — higher percentiles indicate greater predicted risk.
What is the CVSS score of CVE-2026-54234?
CVE-2026-54234 has a CVSS v3 base score of 7.5 (NVD).
How do I remediate CVE-2026-54234?
Patch to the fixed version published by the affected vendor. Where vendor advisories exist for CVE-2026-54234, EchelonGraph cross-links them in the Vendor Advisories panel below — those typically contain the canonical remediation steps, fixed version numbers, and any vendor-specific mitigations.

Dependency Blast Radius

Explore the affected products and dependency analysis for CVE-2026-54234

Explore →

Is Your Infrastructure Affected by CVE-2026-54234?

EchelonGraph automatically scans your cloud infrastructure and maps CVE exposure using blast radius analysis.

Start Free Scan →Browse All CVEs