CVE-2026-53805: NVIDIA SIL GEN3C Unauthenticated RCE via Pickle Deserialization in Inference API

Synopsis

An unauthenticated remote code execution vulnerability exists in NVIDIA Spatial Intelligence Lab's GEN3C inference API server. The /request-inference and /seed-model endpoints deserialize raw HTTP request bodies using Python's pickle.loads() without any authentication or input validation, allowing a network-accessible attacker to supply a crafted pickle payload containing a __reduce__ gadget. Successful exploitation gives the attacker arbitrary code execution as the inference process. A patched-image rebuild at commit db2ffe12ced12ddafcec5e0422ee46ce8520746b is available on HarborGuard for environments running an affected version.

HarborGuard Coverage

Exploit Conditions

• Network reachabilityRequired

  The inference API server must be reachable over the network; an attacker sends the crafted pickle payload directly to the exposed API port without needing local access.

• AuthenticationNot required

  The affected endpoints perform no authentication check, so any client that can reach the API port can deliver a malicious payload.

• Victim interactionNot required

  Exploitation is fully server-side; no user action, click, or session is needed to trigger deserialization.

• Attack complexityDetail

  The exploit is reliable and condition-free; pickle deserialization executes the attacker-supplied __reduce__ gadget deterministically with no race condition or memory-layout dependency.

Blast Radius

• Attacker executes arbitrary operating system commands as the inference process user, gaining an interactive foothold inside the container.
• Attacker reads model weights, inference inputs, and any secrets or credentials mounted into the container filesystem or environment variables.
• Attacker modifies or deletes model artifacts and cached inference state persisted in the container's writable layer or mounted volumes.
• Attacker can crash or hang the inference service, causing a denial of service for any workloads depending on the GEN3C inference API.