GPU VulnDB

Database/AI/ML frameworks & serving

MKP Kubernetes MCP server: unauthenticated log request exhausts server memory

CVSS 7.5CVE-2026-50125AI/ML frameworks & servingcurated

Impact

MKP exposes an unauthenticated get_resource tool on its default HTTP endpoint that accepts caller-controlled limitBytes and tailLines for the pod logs subresource, then copies the whole Kubernetes log stream into an in-memory buffer with no application-side cap. One tools/call request against a pod with large accumulated logs drives huge allocations - the advisory reports over 1 GiB of RSS growth for a 128 MiB requested stream - killing the MKP process. The request-frequency limiter does not constrain per-request volume, so rate limiting does not help. Where MKP is wired into agent tooling that reaches a GPU cluster, this is a denial of that control path, and the same endpoint being unauthenticated by default is the larger operational problem.

Who can reach it

Anyone who can reach the MKP MCP endpoint on its default port. No authentication is required in the default configuration.

What to do

Upgrade to MKP 0.4.1 and restart the server. Until then, keep the endpoint off any network a tenant can reach and put authentication in front of it; the default configuration disables auth.

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.