GLX inspects prompts, responses, retrieved context, and generated tool calls in real time. It brings content safety, AI DLP, direction-aware policy, and enforcement together in one deployable service.
Security with measurable performance.
GLX uses a CPU-first pre-filter for high-confidence matches and applies LLM-based judgment only when deeper reasoning is needed.
Risks unique to generative AI
Prompt injection, jailbreaks, harmful content, exposure of personal or confidential information, source code leaks, corporate policy violations — generative AI brings risks that traditional applications don't have.
The operational cost of adding a guard
Verifying these risks by repeatedly calling another large LLM means extra LLM calls, higher latency, more GPU usage, and higher processing costs — making it hard to scale to high traffic volumes.
Process more requests safely on the same GPU.
GLX doesn't send every request straight to a GPU-based model. Repeated or semantically similar requests are judged first by a CPU-based semantic cache, and only requests that need further analysis go to a lightweight, security-specialized model — a structure that cuts unnecessary GPU inference.
Receive the user request
GLX doesn't automatically forward every request to a GPU-based model.
Repeated or similar requests go to the CPU first
Input that's semantically similar to an already-judged request is handled first by the CPU-based semantic cache, so the same security judgment isn't requested from the GPU again.
New risks are analyzed by a lightweight security model
New or complex requests that the cache can't immediately judge are analyzed by GLX's security-specialized, lightweight fine-tuned model.
After the security decision, on to the LLM app
By combining the semantic cache with the lightweight model, GLX aims to verify more generative AI requests within the same GPU footprint.
Protect every AI interaction— from prompts and responses to RAG context and tool calls.
GLX preserves source and trust context, inspects retrieved documents and tool results, and validates generated tool calls before execution.
Input verification
Analyzes user-entered prompts for harmful requests, prompt injection, jailbreak attempts, and possible sensitive-data or policy violations.
Response verification
Checks the LLM's generated response for personal data, confidential information, source code, harmful content, and other things that shouldn't be exposed, and controls it according to policy.
RAG search-result verification
Verifies the risk in retrieval results during retrieval-augmented generation, so sensitive information or malicious instructions in external or internal documents don't make it into the AI's response.
29 safety categories + 1 artifact scanner. One guardrail.
GLX analyzes structural attacks, harmful content, sensitive-data exposure, and internal policy violations that can occur in a generative AI service, from multiple angles. (29 risk categories · 1 output-artifact scanning, Korean, English, Chinese, and Japanese support)
Prompt Attack
Analyzes structural attacks aimed at disabling a generative AI's instructions — prompt injection, jailbreaks, and instruction bypass.
Harmful Content
Detects content that threatens service operation, including illegal content, harmful content, and dangerous requests.
Sensitive Data
Verifies sensitive data that must not be exposed — personal information, credentials, financial information, health information, and more.
Corporate Assets
Verifies the risk of leaking internal corporate assets — source code, system information, confidential information, business information, and more.
Policy Violation
Controls internal policy violations, disallowed requests, and inappropriate AI responses.
Detect. Decide. Enforce.
Apply pass, replace, or reject actions based on the risk tag, traffic direction, profile, and API key.
User prompt input
The prompt entered by the user is forwarded to GLX.
Analyze for risk
GLX analyzes risk signals such as prompt injection and harmful content in real time.
Block risky requests
Requests found to be high-risk are blocked, along with an explanation of why.
Return a safe response
Returns a safe outcome along with a notice that the request violates safety policy and can't be answered.
Add security, without giving up speed.
Lower Latency
Minimizing the response delay caused by security verification is a core design goal.
Lower GPU Load
Reduces resource load by cutting unnecessary GPU inference calls.
Higher Throughput
Aims for a structure that can handle more requests on the same resources.
From internal chatbots to AI agents and security gateways.
Customer-facing generative AI
Detects malicious or harmful prompts entered directly by customers, and controls inappropriate AI-generated responses and sensitive-data exposure.
Internal generative AI
Reduces the risk of exposing internal sensitive data — personal information, confidential data, source code, system information — entered by employees or output by the AI.
Internal knowledge search & RAG
Verifies that sensitive information and malicious prompts in searched documents and knowledge bases don't make it into the generative AI's response.
AI Agent-based services
Can extend as a Guardrail that controls prompt attacks, privilege bypass, sensitive-data exposure, and policy violations while an AI Agent accesses tools and data.
Add a Guard Layer to your existing AI service.
GLX is designed to connect as a Guard Layer between the user and the LLM app — and, where needed, across the RAG/knowledge-base layer as well.
Add a Guard Layer to your existing AI service
GLX is designed to connect as a Guard Layer between the user request and the AI response, and can be applied the same way to the RAG/knowledge-base layer when needed.
Lightweight, even inside your own environment
Built on a lightweight structure with reduced GPU dependency, GLX is designed to be applied to your internal LLM environment.
Not locked into any single LLM
GLX is designed to integrate as a Guard Layer with your existing generative AI application; check the current officially supported scope separately at the time of adoption.
Controlled safely. Run lightly.
Controls risky input and output
Defends against prompt attacks and blocks harmful content, protects sensitive data, and protects source code and confidential information.
Minimal impact on AI service performance
Aims for lower GPU load and low latency through a lightweight model and semantic cache.
AI usage control aligned to corporate policy
Aims for a Guardrail structure that can apply organization-specific policy based on allow/block rules and risk categories.
Runs light. Proven by security performance.
Security
Verifies detection performance by risk category.
Speed
Verifies latency based on input length and request volume.
Resource
Verifies CPU/GPU resource usage.
Scale
Verifies performance handling concurrent requests and high traffic volumes.
Frequently asked questions about GLX.
The LLM already has its own safety features — why do I need a separate GLX?
An LLM's built-in safety mostly works around the common policy defined by the model provider. GLX sits as a separate Guardrail layer in front of and behind your generative AI application, additionally verifying the risks your specific operating environment needs — prompt attacks, sensitive data, internal corporate policy, and more.
Can it detect prompt injection and jailbreaks?
Yes. GLX analyzes prompt injection and jailbreak attempts — which try to bypass or disable a model's existing instructions — as a major risk category for generative AI.
Does it check for leaks of personal data and corporate confidential information?
GLX is designed to verify the potential exposure of sensitive data — personal information, credentials, source code, system information, business information — in both user input and AI responses.
Won't adding GLX slow down AI response times?
Reducing the latency and resource burden caused by security verification is a core design goal of GLX. Repeated or similar requests are judged first by a CPU-based semantic cache, and only requests needing further analysis go to the lightweight security model — cutting unnecessary GPU inference.
Which LLMs can it integrate with?
GLX aims for a Guardrail structure that isn't locked into any single LLM. That said, the current list of officially supported models and frameworks can vary by product version and deployment method, so check the supported scope at the time of adoption.
Can it be used in an on-premise environment?
GLX was developed with a lightweight structure that reduces dependency on GPU servers, so it can be applied to your internal on-premise LLM environment. Current delivery method and system requirements are confirmed separately based on your deployment environment.
Does it support languages other than Korean?
The latest version of GLX is built to support generative AI security verification in four languages — Korean, English, Chinese, and Japanese.
Start a fast, light, and safe generative AI environment.
Use GLX to verify risk across your generative AI service — from input to response — while cutting the performance overhead of security.
How does this product behave in your environment?
Whether you're exploring, evaluating, or rolling out, you connect directly with a SANDS Lab solutions engineer. Clear every question before contract — that's the point.
Product evaluation & PoC
Real-data PoCs, technical deep-dive sessions, and custom integration scoping. Everything you'd need to validate technical fit before the paperwork starts.
Technical collaboration & licensing
If you want to use the product in an academic benchmark or co-authored paper, we support research licenses and the underlying datasets. Co-authorship is on the table.