If you discover a security vulnerability, please report it responsibly. Do not open a public issue for security problems.

• GitHub Security Advisories: submit a private report.
• Bug Bounty: this project does not currently offer a public bug bounty program.

• A clear and concise description of the vulnerability.
• Steps to reproduce the issue.
• Any potential attack scenario or security impact.
• Suggested mitigations or fixes, if available.

This device plugin registers AMD GPUs to Kubernetes and applies HAMi soft vGPU limits: memory through HIP_DEVICE_MEMORY_LIMIT_<i> and the libamvgpu.so LD_AUDIT hook, plus a kernel dmem cgroup cap where the node supports it, and compute through HSA_CU_MASK. These limits are for cooperative multi-tenant sharing on a trusted cluster. The hook and the CU mask are not a hard security boundary against a workload with enough privilege to bypass its own hook, for example by unsetting the audit library, using a static binary, or ptrace.

• A report that a workload can exceed its own quota, without affecting another tenant, is not a new vulnerability by itself.
• A report that lets a workload reach another tenant's data, device, or namespace it was not granted is in scope; please report it through the process above.

Response times may be affected by weekends, holidays, or time-zone differences. The maintainers aim to reply as soon as possible, ideally within 5 working days.

This project relies on third-party Go modules and container base images. We monitor them with Dependabot and apply security patches promptly.

Thank you for helping keep the project secure.