Artificial intelligence governance startup nFOX Lab today launched a platform for managing owned large language models and announced plans to open-source Blockhouse, a security sandbox designed to restrict what AI workloads can access and do.
The firm, which emerged from stealth this week, aims to help enterprises adopt open models safely on infrastructure they control while limiting the damage a compromised or misbehaving workload could cause. Blockhouse removes privileges by default and lets administrators restore access deliberately through policies enforced at the Linux kernel level.
The company’s approach borrows from zero-trust security principles by sharply limiting access and letting administrators restore privileges selectively. “Every model has the potential to be hostile,” said Gordon Allott, nFOX’s founder and chief executive. “We just presume hostility.”
That addresses a dilemma for companies that want AI capabilities but hesitate to send sensitive information to external providers. Running models internally gives enterprises control over their data, but it also makes them responsible for securing the software and infrastructure that support those models.
Blockhouse operates beneath the application layer, using Linux security mechanisms to restrict model workloads. Rather than relying on instructions telling a model what it should avoid, the system removes capabilities that could enable unauthorized activity, according to Allott.
“You’re basically removing the ability to do it in the first place,” he said.
Communications pass through Porthole, a monitored channel that allows requests to be reviewed before they are fulfilled. Runtime credentials and audit controls remain outside the model process. Administrators can restrict access to files and approve the destinations workloads may contact using a whitelist.
Although the approach is comprehensive, it doesn’t preclude the risk of human error. Allott acknowledged that an administrator could mistakenly authorize a malicious destination.
Blockhouse also checks model integrity. Approved files are sealed with SHA-256 hashes before serving, and built-in checks halt serving if those files change. The software records containment events and reports when a host can’t provide the expected protection.
The sandbox supports the vLLM, SGLang and llama.cpp inference engines. The accompanying nFOX Platform provides a console for evaluating, deploying and monitoring models across cloud virtual machines, rented graphics processing unit hosts and on-premises Linux servers equipped with Nvidia Corp. GPUs.
Teams can check model configurations, verify files and test models with adversarial prompts before deployment. Production workloads receive OpenAI Group PLC-compatible endpoints, while operational evidence can feed existing endpoint detection and security information management systems. Audit records use JavaScript Object Notation and Parquet formats.
Allott said nFOX plans to release Blockhouse under the Apache 2.0 license, while charging an annual license for its fleet management platform. He said open-sourcing the sandbox invites the public scrutiny that security software needs.
“The best security comes from many, many eyes and smart people working on it,” he said.
Allott pointed to Nvidia’s OpenShell as a similar technology that’s focused on securing agents and said nFOX intends to incorporate it. The startup’s roadmap has it expanding into agent protection over time.
The startup company is self-funded but intends to raise venture capital. Allott said nFox has paying beta pilot customers, though it is not yet profitable.
The nFOX Platform is available now. Blockhouse is available in beta test for qualified enterprise evaluations. The commercial platform adds deployment and oversight to the open-source sandbox, with both components focused on keeping privately operated AI under enterprise control.