vLLM bad_words token index validation flaw in SamplingParams
vLLM
A vulnerability in vLLM allows attackers to supply out-of-bounds token indices that corrupt logits memory of concurrent requests. This causes different in-flight HTTP requests to return incorrect tokens.
disclosed CVENo confirmed exploitation, so this record is tracked as a disclosed vulnerability and not counted as an incident. It moves into the counts if CISA or the vendor confirms exploitation. Why
- CVSS
- 3.1
- EPSS
- <0.01 · top 90% · 2026-11-13
- fix status
- Fix status unknown
- first report
- 19 Sep 2026
- last update
- 19 Sep 2026
- CVE ids
Timeline · 1
Sources · 1
Advisories and databases
- 2026-09-19CVE-2026-93989: vLLM through 0.29.0 fails to properly validate bad_words token indices against the model's generation output width in…NVD · nvd.nist.gov/vuln/detail/CVE-2026-93989
Titles and links belong to their publishers. The summary above is written by gcdRadar in its own words; data is CC BY-SA 4.0.
Classification
- lane
- Patch now
- AI role
- Attack on AI
- categories
- Infrastructure vulnerability
- severity
- 2 · Disclosed vulnerability
- fix status
- Fix status unknown
- targets
- Framework / tooling
- owasp llm top 10
- LLM03:2025 Supply Chain
- mitre atlas
- no technique for these categories
- tags
- vllm
- dates
- happened 19 Sep 2026 · first reported 19 Sep 2026
- origin
- NVDVulnerability record from NVD / the GitHub Advisory Database. This product uses the NVD API but is not endorsed or certified by the NVD.
- record
- confidence 0.95 · created 2026-10-18 06:33Z · updated 2026-10-18 06:33Z
Lane, AI role, categories and severity come from the classifier; the OWASP and ATLAS rows are derived from the categories. How these are assigned
Copy a brief
This record as a short Markdown brief, for a ticket, a chat or a coding agent.
the brief as text
# vLLM bad_words token index validation flaw in SamplingParams https://gcdradar.vercel.app/incidents/vllm-bad-words-token-index-validation-flaw-in-samplingparams · AIR-2026-0152 · closed 2026-09-19 - Lane: Patch now - Severity: 2 (Disclosed vulnerability) - Fix status: Fix status unknown - First reported: 2026-09-19 - AI role: Attack on AI - Products: vLLM - CVE ids: CVE-2026-93989 - CVSS: 3.1 ## Summary A vulnerability in vLLM allows attackers to supply out-of-bounds token indices that corrupt logits memory of concurrent requests. This causes different in-flight HTTP requests to return incorrect tokens. ## What to do No action recorded; read the sources below and the vendor's advisory. ## Latest updates - 2026-09-19: CVE-2026-93989: vLLM through 0.29.0 fails to properly validate bad_words token indices against the model's generation output width in… (https://nvd.nist.gov/vuln/detail/CVE-2026-93989) ## Sources - 2026-09-19 NVD: CVE-2026-93989: vLLM through 0.29.0 fails to properly validate bad_words token indices against the model's generation output width in… https://nvd.nist.gov/vuln/detail/CVE-2026-93989 Verify against the vendor's advisory before changing production. Data CC BY-SA 4.0, https://gcdradar.vercel.app.
Related
- closed7.1vLLM CVE-2026-109311: vLLM audio transcription endpoint heap overflow vulnerability
Patch now · Disclosed · since 12 Oct 2026 · CVE-2026-109311
AI - closed4.3vLLM CVE-2026-105775: Out-of-bounds read vulnerability in vLLM completions request handler
Patch now · Disclosed · since 6 Oct 2026 · CVE-2026-105775
AI - closed4.3vLLM Denial of Service Vulnerability in vLLM Penalty Handler
Patch now · Disclosed · since 6 Oct 2026 · CVE-2026-105922
AI - closed5.3vLLM vLLM Denial of Service via Unbounded Video Sampler Frames
Patch now · Disclosed · since 5 Oct 2026 · CVE-2026-105758
AI - closed3.1vLLM vLLM tool continuations drop cache_salt to restore prefix-cache oracle
Patch now · Disclosed · since 5 Oct 2026 · updated 6 Oct 2026 · CVE-2026-105752
AI - closed5.3vLLM vLLM GLMGA video sampling permits CPU and memory exhaustion
Patch now · Disclosed · since 5 Oct 2026 · CVE-2026-105760
AI