nFOX Lab to open-source Blockhouse sandbox for secure AI model deployment
AI governance startup nFOX Lab launched a platform for managing owned large language models and plans to open-source its security sandbox, Blockhouse, to help enterprises restrict AI workload access.
Key points
- nFOX Lab emerged from stealth and launched a platform for managing owned large language models on enterprise-controlled infrastructure.
- The company plans to open-source Blockhouse, a security sandbox that restricts AI workload capabilities using Linux kernel-level policies.
- Blockhouse supports vLLM, SGLang, and llama.cpp inference engines and uses SHA-256 hashes to verify model integrity before serving.
Artificial intelligence governance startup nFOX Lab announced today that it will open-source Blockhouse, a security sandbox designed to restrict the actions of AI workloads. The firm, which emerged from stealth this week, also launched a broader platform for managing large language models that enterprises own and operate on their own infrastructure. This move aims to help organizations adopt open models safely while limiting the potential damage from compromised or misbehaving software.
Blockhouse operates beneath the application layer, utilizing Linux security mechanisms to enforce strict access controls. According to nFOX, the system removes privileges by default rather than relying on instructions that tell a model what to avoid. Administrators can deliberately restore specific access through policies enforced at the Linux kernel level, adhering to zero-trust security principles that sharply limit initial permissions.
Security Architecture and Integrity Checks
The sandbox architecture addresses the dilemma faced by companies that want AI capabilities but hesitate to send sensitive data to external providers. Running models internally gives enterprises control over their data, but it also places the responsibility for securing the supporting software and infrastructure on the organization. Blockhouse mitigates this by removing capabilities that could enable unauthorized activity before they can be exploited.
Gordon Allott, nFOX’s founder and chief executive, stated that every model has the potential to be hostile, so the company presumes hostility in its design. Communications from the model pass through Porthole, a monitored channel that allows requests to be reviewed before fulfillment. Runtime credentials and audit controls remain outside the model process, ensuring that sensitive information is not exposed to the AI workload directly.
Administrators can restrict access to files and approve the specific destinations that workloads may contact using a whitelist system. Although the approach is comprehensive, Allott acknowledged that human error remains a risk, noting that an administrator could mistakenly authorize a malicious destination. The system also checks model integrity by sealing approved files with SHA-256 hashes before serving them.
Supported Engines and Monitoring
Blockhouse supports several popular inference engines, including vLLM, SGLang, and llama.cpp, allowing for broad compatibility with existing AI infrastructure. The software includes built-in checks that halt serving if the sealed files change, ensuring that the model has not been tampered with since approval. This integrity verification is critical for maintaining trust in the output of large language models running in sensitive environments.
The platform records containment events and reports instances where a host cannot provide the expected protection levels. This logging capability allows security teams to audit the behavior of AI workloads and identify potential policy violations or configuration errors. By keeping runtime credentials separate from the model process, the system reduces the attack surface available to any compromised workload.
What should you do?
Organizations evaluating open-source AI models should assess their current infrastructure for similar sandboxing capabilities. Security teams should review how they manage runtime credentials and ensure that AI workloads do not have unnecessary access to sensitive files or network destinations. Monitoring for unauthorized communication attempts from AI processes is essential for detecting misconfigurations or hostile model behavior.
Background: Large language models
Large language models are a specific type of neural network trained on text. Generative AI is the broader umbrella term for any system that creates new content, including images and code. Not all generative AI uses large language models, but most modern text generators do.
What to do and how to stay safe: nFOX Lab
- Review current AI workload permissions to ensure they operate with minimal privileges and do not have direct access to sensitive credentials.
- Implement integrity checks for AI model files using cryptographic hashes to detect any unauthorized modifications before deployment.
- Monitor network traffic from AI inference engines to identify and block connections to unapproved destinations or external services.
- Audit administrative policies regularly to ensure that whitelisted destinations and file access rules remain accurate and free from human error.
Step-by-step guide: Prompt Injection Warning Signs: How to Spot AI Manipulation
General security guidance from the Threat Times newsroom. It is not confirmed advice from the organisations named in this story.
Frequently asked questions
What inference engines does Blockhouse support?
Blockhouse supports the vLLM, SGLang, and llama.cpp inference engines, according to nFOX Lab.
How does Blockhouse handle model integrity?
Approved files are sealed with SHA-256 hashes before serving, and built-in checks halt serving if those files change.
Is Blockhouse currently open-source?
nFOX Lab announced plans to open-source Blockhouse, but the material does not state that the code is publicly available yet.
