nFox has launched a platform for managing AI models run on companies’ own infrastructure, and says it plans to release its security sandbox as open source. The practical pitch is to limit what a model can access or do, rather than relying on the model to behave itself.
Watch Desk analysis
What happened
SiliconANGLE reports that the nFOX Platform is available now, while Blockhouse, the sandbox, is in beta for qualified enterprise evaluations. The company plans to release Blockhouse under the Apache 2.0 licence. It supports vLLM, SGLang and llama.cpp.
According to the report, Blockhouse uses Linux kernel-level controls to restrict workloads by default. Administrators can set file access and approved network destinations, while a monitored channel called Porthole reviews requests. Model files are checked against SHA-256 hashes, with serving halted if they change. The wider platform offers tools for evaluating and monitoring models across cloud and on-premises Linux systems.
Why it matters
Running models on internal infrastructure gives organisations more control over where data goes, but makes them responsible for the surrounding security. A sandbox that limits a model’s permissions at the operating-system level addresses a different problem from asking it, politely and repeatedly, not to do anything alarming.
The report says nFox’s platform is live, but Blockhouse remains in a limited beta and the open-source release is planned, not yet delivered. The company’s founder also acknowledged that an administrator could mistakenly approve a malicious destination. Those details matter: controls can narrow the blast radius, but they do not make configuration mistakes disappear.
Our read
This is a useful AI infrastructure development because it puts concrete controls around models that organisations operate themselves. The open-source plan could make the sandbox easier to scrutinise, but the meaningful test will be the code, its documentation and how well the controls work in real deployments. For now, treat this as a promising security approach with a beta product, not a proven shield.
What to watch
- When Blockhouse is released under Apache 2.0, and what the public code contains.
- Whether nFox publishes technical details or independent evaluations of the sandbox.
- How the platform handles permissions, audit records and mistaken administrator approvals in customer deployments.
Discussion spark: For AI systems running inside organisations, should security start with strict operating-system limits on what models can do, or with better testing and oversight of the models themselves?
Sources and evidence
- Startup nFox to open-source its 'sandbox' for open AI models – SiliconANGLE (15 October 2026, 15:54 UTC)
Watch Desk is operated by WittyWires as an independent cross-cutting AI news tracker. It does not speak for the organisations or people it covers.