Tools AI Risk Radar ai-incident-0091
Incident record
One negative token ID crashes the GPU behind vLLM's embeddings routes until a restart
- Severity
- High
- Status
- Patched
- Type
- Infra Vuln
- Target
- vLLM before 0.28.0, /v1/embeddings and /pooling endpoints
- Actor
- researcher
- CVE
- CVE-2026-93592
What happened
vLLM validated only the upper bound of token IDs on the embeddings and pooling routes, so one unauthenticated request carrying a negative token ID triggers a CUDA device-side assertion that takes down the GPU context until the server restarts. Scored CVSS 3.1 7.5 and fixed in 0.28.0.
The advisory published 3 September; the CVE record reached NVD on 18 September through VulnCheck, a paperwork catch-up on an already-patched flaw.
Sources
- vLLM advisory GHSA-25q3-v2hm-8vpfgithub.com/vllm-project/vllm/security/advisories/GHSA-25q3-v…
- NVD record CVE-2026-93592nvd.nist.gov/vuln/detail/CVE-2026-93592
One record from the AI Risk Radar, maintained by the Security Desk. Data: CSV · JSON ·RSS · CC BY 4.0 with attribution to YFarmX.Tracker updated · 18 September 2026