GPU VulnDB

Database/AI/ML frameworks & serving

vLLM Rust gRPC frontend: >128 output_candidates.token_ids triggers a fatal sampler exception

CVSS 8.2CVE-2026-116173AI/ML frameworks & servingcurated

Impact

The Rust gRPC frontend does not enforce the logprob_token_ids limit, so a Generate or GenerateStream request with output_logprobs enabled and more than 128 output_candidates.token_ids raises a fatal exception in the sampler and terminates EngineCore. The replica stops serving and all co-tenant requests in flight are lost; recovery requires the process to restart and reload weights onto the GPUs. This is one of three distinct unvalidated-field crashes in the same gRPC frontend (CVE-2026-116172, CVE-2026-116174) fixed in the same release.

Who can reach it

Network, unauthenticated - any client able to reach the vLLM Rust gRPC Generate or GenerateStream service.

What to do

Upgrade to vLLM 0.24.0 or later and restart the serving processes - rolling restart of inference replicas, no node drain. Interim: keep the gRPC frontend off untrusted networks and cap the number of candidate token ids at the proxy.

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.