Database/AI/ML frameworks & serving
NVIDIA OpenShell: sandbox status metadata steers the gateway into connecting to arbitrary internal hosts (SSRF)
Impact
The OpenShell control-plane gateway takes its connection target from a sandbox's status metadata and does not check it against the expected address range. An attacker who can edit that metadata can make the gateway connect to any internal host it can reach, which is server-side request forgery from the control plane. Confidentiality impact is high (scope changed). Exploitation needs high privileges and high complexity.
Who can reach it
Remote attacker over the network with high privileges, able to change a sandbox's status metadata. No user interaction needed.
What to do
Update NVIDIA OpenShell to the fixed release in NVIDIA product-security bulletin 5929. The record does not give a fixed version. Until then, limit what the gateway can reach on the network and who can write sandbox status metadata.
References
Related entries
- llama.cpp (GGUF vocabulary parsing, llama_vocab::impl::print_info): MALICIOUS MODEL FILE CRASHES THE SERVER: the GGUFNCVD-2025-020-llama-cpp-gguf-vocabulary-parsin · llama.cpp (GGUF vocabulary parsing, llama_vocab::impl::print_info)Medium
- Kubeflow: SSRFCVE-2023-6570 · KubeflowMedium
- Gradio: SSRF in the `/proxy` routeCVE-2024-2206 · GradioMedium
- picklescan: ZIP manipulation crashes the scanner (scan bypass by DoS)CVE-2025-1944 · picklescanMedium
- vLLM (`/v1/completions` guided decoding): Invalid `json_schema` kills the serverCVE-2025-48942 · vLLM (`/v1/completions` guided decoding)Medium
- NVIDIA OpenShell: unbounded chunked requests exhaust the inference-interception parserCVE-2026-100982 · NVIDIA OpenShell for Linux (inference-interception request parser)Medium
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.