CVE-2026-93989

MEDIUMNVD 4.34.3—
EchelonGraph scoreMEDIUM confidence

This medium-severity CVE scores 4.3 under NVD CVSS v3. EPSS exploit probability: 0.2% (9th percentile of EPSS-scored CVEs). GitHub Security Advisory enrichment pending alignment with NVD CVSS.

Triggered by: NVD CVSS baseline
Sources: epss, ghsa, nvd
Trending — 4 sources updated this week
4.3EG
EchelonGraph verdictMonitorLow exploitation likelihood right now — keep watching.
  • Lower severity and no public exploit yet
CISA-KEV: Not listedEPSS PROB: 0.2%CVSS: 4.3Exploit: None knownExposed services: Not assessed

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

vLLM through 0.29.0 fails to properly validate bad_words token indices against the model's generation output width in SamplingParams.update_from_tokenizer(). Attackers can supply out-of-bounds token indices that corrupt logits memory of concurrent requests, causing different in-flight HTTP requests to return incorrect tokens.

CVSS v3
4.3
EG Score
4.3MEDIUMmedium confidence
EG Risk
19
EG Risk 19/100CISA SSVC

EG Risk is EchelonGraph's 0–100 priority score: it fuses intrinsic severity with real-world exploitation and automatability so you can rank equal-severity CVEs and fix the most dangerous first. Higher = act sooner. Distinct from the 0–10 EG Score (severity).

How it’s computed
Severity43% × 45%
Exploitation0% × 40%
Automatability0% × 15%
CISA SSVC: Track at every mission impact level.
Action: No fix is confirmed yet. Restrict network exposure of the affected system or apply the vendor's mitigation within your standard update timelines, and watch the vendor's advisory for the fix.
EPSS PROB
0.2%
EPSS %ILE
9th
KEV
Not listed

CISA SSVCTrack at every mission impact level.

No fix is confirmed yet. Restrict network exposure of the affected system or apply the vendor's mitigation within your standard update timelines, and watch the vendor's advisory for the fix.

Exploitation none (CISA Vulnrichment) · Automatable no (CISA Vulnrichment) · Technical impact partial (CISA Vulnrichment). Mission impact is CISA's Mission & Well-being decision point, and only you can judge it: high means the affected system is essential to your organisation's mission, or its compromise could cause irreversible harm to people. CISA's decision table

Published

September 19, 2026

Last Modified

September 28, 2026

Advisory Details (5)

Auto-updated Sep 28, 2026
No patch confirmed yet. Sources: github_pr.
generic

vLLM through 0.29.0 Cross-Request Logits Corruption via bad_words | Advisories | VulnCheck

https://www.vulncheck.com/advisories/vllm-through-0.29.0-cross-request-logits-corruption-via-bad-words
github_pr

fix(security): validate bad_words token IDs against model vocab size

Upstream fix in progress (PR #48824 · vllm-project/vllm, state=open)

https://github.com/vllm-project/vllm/pull/48824
generic

vllm/vllm/v1/worker/gpu/sample/bad_words.py at 98dff2a81d747d1dba01a47f939f48c3526d4206 · vllm-project/vllm · GitHub

https://github.com/vllm-project/vllm/blob/98dff2a81d747d1dba01a47f939f48c3526d4206/vllm/v1/worker/gpu/sample/bad_words.py
generic

vllm/vllm/sampling_params.py at 98dff2a81d747d1dba01a47f939f48c3526d4206 · vllm-project/vllm · GitHub

https://github.com/vllm-project/vllm/blob/98dff2a81d747d1dba01a47f939f48c3526d4206/vllm/sampling_params.py#L694-L753
generic

GitHub - vllm-project/vllm: A high-throughput and memory-efficient inference and serving engine for LLMs · GitHub

https://github.com/vllm-project/vllm

Vendor Advisories for CVE-2026-93989(1)

These vendors published their own advisory mentioning this CVE — often with vendor-specific remediation steps + affected product lists not in NVD.

Affected Packages

(1 across 1 ecosystem)
PyPI(1)
PackageVulnerable rangeFixed inDependents
vllm0.0.1 ... 0.9.2 (97 versions)0.30.0—

Weakness Classification(1)

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

Data Freshness Timeline

(refreshed 20× in last 7d / 36× 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. 2026-10-04 09:03 UTCEG score recompute
  2. 2026-10-04 09:03 UTCGHSA enrichment
  3. 2026-10-03 14:26 UTCEPSS rescore
  4. 2026-10-03 06:53 UTCEG score recompute
  5. 2026-10-03 06:53 UTCGHSA enrichment
  6. 2026-10-02 17:57 UTCEPSS rescore
  7. 2026-10-02 04:43 UTCEG score recompute
  8. 2026-10-02 04:43 UTCGHSA enrichment
  9. 2026-10-01 19:51 UTCEPSS rescore
  10. 2026-10-01 02:33 UTCEG score recompute
  11. 2026-10-01 02:33 UTCGHSA enrichment
  12. 2026-09-30 15:04 UTCEPSS rescore
  13. 2026-09-30 00:23 UTCEG score recompute
  14. 2026-09-30 00:23 UTCGHSA enrichment
  15. 2026-09-28 18:40 UTCEG score recompute▲ 1.20
  16. 2026-09-28 18:40 UTCGHSA enrichment
  17. 2026-09-28 18:40 UTCNVD updateCVSS v3 → 4.3 · severity → MEDIUM
  18. 2026-09-28 13:52 UTCEPSS rescore
  19. 2026-09-28 13:52 UTCEPSS rescore
  20. 2026-09-27 13:49 UTCEPSS rescore
  21. 2026-09-26 15:59 UTCEPSS rescore
  22. 2026-09-26 00:58 UTCEG score recompute
  23. 2026-09-26 00:57 UTCGHSA enrichment
  24. 2026-09-24 14:04 UTCEPSS rescore
  25. 2026-09-23 17:54 UTCEPSS rescore
Show 11 more
  1. 2026-09-22 20:32 UTCGHSA enrichment
  2. 2026-09-22 16:22 UTCEG score recompute
  3. 2026-09-22 16:22 UTCGHSA enrichment
  4. 2026-09-22 16:06 UTCEG score recompute
  5. 2026-09-22 16:06 UTCGHSA enrichment
  6. 2026-09-22 16:01 UTCEPSS rescore
  7. 2026-09-21 21:09 UTCEPSS rescore
  8. 2026-09-20 20:16 UTCEPSS rescore
  9. 2026-09-19 23:27 UTCEG score recompute
  10. 2026-09-19 23:00 UTCEG score recompute
  11. 2026-09-19 23:00 UTCMITRE cvelistV5first tracked

Frequently asked(5)

What is CVE-2026-93989?
CVE-2026-93989 is a medium vulnerability published on September 19, 2026. vLLM through 0.29.0 fails to properly validate badwords token indices against the model's generation output width in SamplingParams.updatefrom_tokenizer(). Attackers can supply out-of-bounds token indices that corrupt logits memory of concurrent requests, causing different in-flight HTTP requests…
When was CVE-2026-93989 disclosed?
CVE-2026-93989 was first published on September 19, 2026, with the most recent update on September 28, 2026. EchelonGraph re-ingests CVE updates from NVD on a 2-hour cycle, so this page reflects the latest published state.
Is CVE-2026-93989 actively exploited?
CVE-2026-93989 is not currently on CISA's Known Exploited Vulnerabilities catalog. FIRST EPSS estimates a 0.2% probability of exploitation in the next 30 days (9th percentile of EPSS-scored CVEs).
What is the CVSS score of CVE-2026-93989?
CVE-2026-93989 has a CVSS v3 base score of 4.3 (NVD).
How do I remediate CVE-2026-93989?
No fix for CVE-2026-93989 is confirmed yet. Until one is published, restrict network exposure of the affected system or apply the vendor's mitigation — for example, keep it off the internet or limit it to trusted networks — and watch the vendor's advisory for the fix. The vendor advisories EchelonGraph has for CVE-2026-93989 are linked in the Vendor Advisories panel on this page.

Dependency Blast Radius

See which npm, PyPI, Go, and Maven packages are affected by CVE-2026-93989

Explore →

Is Your Infrastructure Affected by CVE-2026-93989?

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