A reachable inference port on a GPU host becomes code execution when the service treats the HTTP body as a Python object. The standard response would focus on patching the bug, but the real break is the trust boundary: any client that can reach the API can make the inference process run attacker code.
VulnCheck says NVIDIA GEN3C's unauthenticated RCE, tracked as CVE-2026-53805, sits in two FastAPI endpoints that pass raw request bodies to `pickle.loads()` with no authentication, validation, or restricted unpickler. The project listens on port 8000 by default and is commonly reached over an SSH tunnel, so the exposed surface is a remote GPU host in lab or cloud deployments; NVIDIA fixed the deserialization path at commit `db2ffe1`, and earlier checkouts remain exposed.
The broader risk is not just this one project. VulnCheck found GEN3C through a GitHub search for public AI/ML code that routes raw request bodies into `pickle.loads()`, which means the same pickle-over-HTTP pattern can recur in other inference services.