Executive brief
NVIDIA GEN3C is a research project from the Spatial Intelligence Lab used for 3D generative modeling. A security flaw in its inference server allows an unauthenticated attacker to remotely execute commands on the host system. This could lead to a complete takeover of the server, theft of proprietary AI models, or unauthorized access to sensitive data processed by the application.
Technical details
The NVIDIA Spatial Intelligence Lab (SIL) GEN3C inference API server is vulnerable to unauthenticated remote code execution (RCE). The root cause is the use of Python's 'pickle.loads()' to deserialize raw HTTP request bodies at the '/request-inference' and '/seed-model' endpoints without prior authentication or input validation. An attacker can exploit this by sending a specially crafted HTTP POST request containing a malicious '__reduce__' gadget. Successful exploitation allows the attacker to execute arbitrary code with the privileges of the inference process. The vulnerability was addressed by replacing pickle-based deserialization with a safer 'safetensors' and JSON-based codec in commit db2ffe1.
Affected products
- NVIDIA GEN3C versions before commit db2ffe12ced12ddafcec5e0422ee46ce8520746b
Timeline
- 2026-06-11: disclosed: Initial fix proposed via community pull request
- 2026-06-15: patched: Official patch committed to main branch
- 2026-06-17: advisory: NVD and VulnCheck advisories published