Tracking: CVE-2026-50125
Severity: High
Who is affected
MKP before 0.4.1, where an attacker can reach its default HTTP interface and the tool can access sufficiently large pod logs. The reported get_resource flow does not require authentication.
How the attack works
The pod-log operation accepts caller-controlled limitBytes and tailLines values without an effective application-side cap. Large log responses are copied into memory, including buffering and subsequent processing. Repeated or oversized requests can exhaust the MCP process's memory.
The result is service degradation or termination, denying legitimate Kubernetes tool access. A Kubernetes permission that allows log reading does not itself impose a safe response-size budget.
What to do
- Upgrade to 0.4.1 or later and restart MKP.
- Restrict the MCP HTTP endpoint to authenticated, trusted clients and enforce reasonable log-size and request limits.
- Review oversized log requests, memory spikes, and process restarts. Limit Kubernetes credentials to the namespaces and resources the tool actually needs.
Scope and limits
Rate limits and container memory limits can reduce the impact while patching, but they do not replace bounds on the data read for a single request.