Orion soft has launched StarGuard AI, a specialized security gateway (AI Firewall) that operates as a reverse proxy between users and language models. The platform is designed to protect corporate data when using both cloud-based and local LLMs.
What Happened
Orion soft's StarGuard AI solution provides centralized access management for models such as ChatGPT, Claude, and DeepSeek, as well as local solutions like GigaChat and YandexGPT. The system implements multi-layered data verification using regular expressions, ML detectors for masking personally identifiable information (PII), and specialized LLM detectors to combat prompt injections, jailbreak attacks, and toxic content. The platform also supports OCR for analyzing attachment content in PDF, XLSX, and DOCX formats and integrates with OpenWebUI, IDEs, and AI agents.
Context
The "Shadow AI" problem arises when employees use third-party neural network services without the knowledge of the security department, creating risks of intellectual property and personal data leaks through prompts. StarGuard AI acts as an infrastructural protective layer, allowing companies to legalize the use of powerful cloud models within a controlled perimeter.
Why It Matters for the Industry
The emergence of specialized AI gateways marks a transition from uncontrolled LLM usage to structured implementation through compliance layers. In the near future, growth in the AI Firewall segment and the standardization of interaction protocols between corporate systems and external model APIs are expected, making such protective layers a mandatory component of enterprise infrastructure.
Why It Matters for Users
For companies, this provides the opportunity to safely use advanced models like Claude or ChatGPT without fear that secret code or data will end up in providers' training sets. Developers gain a tool for creating secure AI agents, with built-in protection against malicious instructions within files and prompts.
What Is Not Yet Known / Limitations
There are technical challenges related to the platform's potential impact on system latency and questions regarding the reliability of the ML detectors themselves in real production environments.
Sources
Author
Look at AI, Editorial Team