Startup nFox to open-source its ‘sandbox’ for open AI models
SiliconANGLE Paul Gillin
nFOX is opening its AI security sandbox, Blockhouse, so companies can run open models on their own gear. The twist: it strips privileges by default, because the company assumes models can go rogue.
Based on reporting by SiliconANGLE, Paul Gillin — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
nFOX Lab has come out of stealth with a simple pitch: if companies want the flexibility of open AI models, they shouldn’t have to trust those models with the keys to the kingdom. To back that up, the startup is planning to open-source Blockhouse, a security sandbox built to tightly limit what model workloads can touch and do.
The core idea is blunt. Blockhouse starts by removing privileges, then gives administrators a way to add access back through policies enforced at the Linux kernel level. It sits below the application layer, uses Linux security controls, and routes communication through a monitored channel called Porthole so requests can be reviewed before they’re carried out. Runtime credentials and audit controls stay outside the model process.
That matters because the whole point of running models internally is control. Enterprises keep their data on systems they own, but they also inherit the security burden that comes with it. Gordon Allott, nFOX’s founder and chief executive, says the company assumes every model could be hostile. So instead of trying to coach a model into behaving, Blockhouse simply removes the ability to carry out risky actions in the first place.
The sandbox also watches for tampering. Approved files are sealed with SHA-256 hashes before serving, and if those files change, serving stops. Administrators can whitelist destinations, restrict file access, and the system records containment events when a host can’t provide the protection it is supposed to. Allott did acknowledge one obvious weak spot: a human can still approve a bad destination by mistake.
nFOX says Blockhouse supports the vLLM, SGLang and llama.cpp inference engines, while its nFOX Platform provides a console for evaluating, deploying and monitoring models across cloud virtual machines, rented GPU hosts and on-premises Linux servers with Nvidia GPUs. The company says the platform can produce OpenAI Group PLC-compatible endpoints and export audit records in JSON and Parquet. Blockhouse is in beta for qualified enterprise evaluations, the platform is available now, and the company plans to release Blockhouse under Apache 2.0 while charging an annual license for fleet management.
My take — AI-written commentary, not fact-checked reporting
This is the right instinct: treat model workloads like untrusted software, because that’s what they are. The industry has spent plenty of time pretending prompt rules are a security strategy; that’s cute, but not serious. Open models are only going to win in enterprises if the controls are real, boring and hard to bypass.
Read more about this at: SiliconANGLE
