StarGuard AI: зачем корпоративным LLM нужен firewall и как его строит Orion Soft

OrionSoft announced StarGuard AI in July 2026 — a security gateway (firewall) for large language models in a corporate environment. The product sits between users and models, controlling how employees, developers, and internal services access local and cloud LLMs.

Why business needed a firewall for LLMs

Over the past year, large language models have made the enterprise journey from isolated experiments to being embedded in everyday workflows — and this is exactly what created a new risk surface. While a model was one team's toy, there was nothing to protect; once an LLM became part of the logic of internal products, every request to it turned into a channel through which data could leak or a malicious instruction could arrive.

According to OrionSoft's description, over the past year companies have simultaneously seen four ways of using LLMs grow at once:

• IT teams deploy local (on-premise) models
• Business users work through cloud services
• Developers plug AI agents directly into the IDE
• Internal products use LLMs as part of their business logic

Each of these four channels lives by its own access rules. Without a single point of control, the security department cannot see who is sending what, and where, into the model.

What the StarGuard AI gateway protects

StarGuard AI works as a single point of control for all requests to LLMs — similar to a network firewall, but for traffic to models. The idea is to give corporate security one managed gateway instead of dozens of separate integrations with cloud or local services.

The typical threats such gateways are built to address lie in three areas: leakage of sensitive data when sending requests to external cloud LLMs, prompt injection — mixing malicious instructions into user input, and "shadow AI," where employees use models in violation of company policy. A single gateway makes it possible to see and filter this flow at one point.

"This is exactly the task

StarGuard AI is built for — a security gateway for large language models," — Nikita Vekesser, product manager for AI infrastructure, OrionSoft.

As Vekesser notes in OrionSoft's blog on Habr, the shift from experimentation to integration was "short but noticeable": the technology was adopted faster than companies managed to build a security perimeter around it. The gateway closes exactly this gap — between the speed of LLM adoption and the maturity of protection.

How an LLM firewall differs from a network one

An LLM firewall does not control network packets, but the meaning of the model's requests and responses — it looks at the content of prompts and generations, not just IP addresses and ports. An ordinary network firewall does not understand that a request to an LLM contains an attempt to extract a corporate secret or bypass an instruction, while a specialized gateway is designed specifically for these four scenarios of corporate model use.

This reflects a broader shift: a separate infrastructure layer is starting to form around language models — access control, request auditing, input and output filtering. Previously such a layer was built around networks, databases, and APIs; now it is the LLM's turn.

What this means

AI infrastructure is entering a phase where a separate security perimeter is growing up around language models — just as it once did around networks and APIs. The emergence of specialized LLM firewalls like StarGuard AI means that corporate use of AI is ceasing to be an experiment and starting to require the same controls as any critical IT system.