NVIDIA Spatial Intelligence Lab's GEN3C contains an unauthenticated remote code execution flaw in its inference API. The server deserializes the raw HTTP body of /request‑inference and /seed‑model calls via Python's pickle.loads() without authentication or input validation. An attacker can send a crafted pickle payload that exploits the __reduce__ gadget, allowing arbitrary code to be executed within the inference process. This flaw falls under CWE‑502 and presents a severe threat to confidentiality, integrity, and availability of the host system.

Affected systems are deployments of NVIDIA SIL GEN3C (nv‑tlabs:GEN3C). No specific version range is supplied in the advisory, so any instance of the product with the vulnerable inference API implementation is considered at risk. The vulnerability is tied to the inference API endpoints exposed on the service.

The CVSS score of 9.3 indicates critical severity, while the EPSS score of <1% shows that, at present, the likelihood of exploitation is low but still possible. The vulnerability is not listed in CISA KEV. The attack vector is remote over the network: an unauthenticated user may reach the exposed inference API port, craft a malicious pickle payload, and trigger code execution with the privileges of the inference process. Successful exploitation would grant the attacker control over the affected host.