GPU VulnDB

Database/AI/ML frameworks & serving

vLLM: malformed tiktoken vocab file crashes the tokenizer backend, denying service on the GPU node

CVSS 1.9CVE-2026-90713AI/ML frameworks & servingcurated

Impact

A local user who can place or point vLLM at a crafted tiktoken vocabulary file triggers a failure in TiktokenTokenizer::new in the Rust text backend, taking the server down. On a GPU node that means the inference process dies and the GPUs it held go idle until the serving pod restarts; no data disclosure or integrity loss is claimed by the record. Severity is rated very low (CVSS 1.9) because it needs local access and an already-privileged path to supply the vocab file. The record notes a public proof of concept exists and the fix is still an unmerged pull request.

Who can reach it

Local user on the serving host with an existing low-privilege account who can influence the tokenizer vocab file vLLM loads. No remote or network path is described.

What to do

No released fix at the time of the record - the upstream pull request is still awaiting acceptance. Until it lands, restrict who can write model and tokenizer directories mounted into vLLM pods and load vocab files only from trusted artifact storage. Applying the fix when released is a container image update and a restart of the vLLM process, not a node reboot.

References

Related entries

All AI/ML frameworks & serving entries

This entry is curated: imported from vendor advisories with machine assistance, not yet individually verified. Confirm against your vendor's advisory before acting, and report anything wrong.