> ## Documentation Index
> Fetch the complete documentation index at: https://docs.beyondguard.ai/llms.txt
> Use this file to discover all available pages before exploring further.
# How BeyondGuard Secures Every Layer of Your AI Stack
> BeyondGuard sits independently between your users and your AI models, inspecting every interaction at every layer without modifying model behavior.
BeyondGuard is positioned inside your AI value chain — not beside it. Rather than observing traffic after the fact or relying on model-side safety mechanisms, BeyondGuard intercepts interactions at the precise points in the stack where threats arise: when a prompt arrives, when an agent forms a plan, when a tool call is executed, when context is retrieved from a vector store, and when a model output is about to be delivered to a user. At every one of these points, BeyondGuard evaluates the interaction against your active guards and policies and takes action before the interaction completes. This architecture means that BeyondGuard's protection is not a post-processing filter — it is a structural part of how your AI system operates.
## The Six Layers
AI systems are not monolithic. A modern enterprise AI deployment typically spans multiple distinct layers, each of which introduces its own threat surface. BeyondGuard covers all six.
Protects embedding pipelines, vector stores, and RAG retrieval systems. BeyondGuard inspects documents and retrieved context for poisoned content, PII-bearing embeddings, and unauthorized data that should not be injected into model context.
Sits between application logic and the foundation model. BeyondGuard ensures that only cleared, policy-compliant inputs reach the model and that the model's responses are evaluated before propagating downstream.
Monitors the application logic that orchestrates model calls. BeyondGuard checks that application-level behavior — including session context and conversation history — does not create pathways for scope violations or data leakage.
Tracks the multi-step execution plans of autonomous agents. BeyondGuard evaluates each step in an agent's plan for deviation from its intended scope, unauthorized tool selection, and memory poisoning that could corrupt future decisions.
Validates every external tool call and Model Context Protocol integration. BeyondGuard checks parameters against expected schemas, flags anomalous call patterns, and blocks tool poisoning attempts before they execute.
Governs the final delivery of model outputs to end users. BeyondGuard reviews responses for toxic content, confidential data leakage, PII exposure, and outputs that would violate regulatory requirements before they are surfaced in the user interface.
## Structural Independence
The most important architectural property of BeyondGuard is that it makes security decisions **outside** the models it protects. This is not an implementation detail — it is a foundational design requirement.
Models can be manipulated through their inputs. A sufficiently crafted prompt can override a model's system instructions, cause it to ignore safety fine-tuning, or trick it into reasoning its way around a guardrail. If your security logic runs inside the model's inference context — as system prompt instructions, for example — then your security logic is subject to the same attacks as the model itself.
BeyondGuard avoids this problem entirely. Its evaluation pipeline is a separate computational process that the model cannot influence, rewrite, or reason around. When BeyondGuard evaluates a prompt for injection, it does so in an isolated context that is structurally inaccessible to the model being protected. When BeyondGuard reviews an agent plan for scope deviation, it applies its own policy logic, not the agent's. This separation holds at every layer.
Structural independence also means that BeyondGuard's security posture does not degrade as you update, fine-tune, or swap out your underlying models. Your guards and policies remain in force regardless of what model version is running beneath them.
## The Control Plane
All five guards — Prompt Guard, Agent Guard, MCP Guard, Data Guard, and Output Guard — report to a single Control Plane. The Control Plane is your unified interface for configuring policies, reviewing threat events, managing projects, and tracking your organization's security maturity level.
Key Control Plane capabilities include:
* **Unified policy management** — Define what each guard does when it detects a threat (block, flag, redact, alert) and apply those policies consistently across all projects and environments from one place.
* **Threat Events dashboard** — Every detected threat is logged with a category, confidence score, guard attribution, and the action taken. You can filter, search, and export events for audit and incident response purposes.
* **Maturity tracking** — The Control Plane tracks your current maturity level across the L0–L5 progression and surfaces recommendations for advancing to the next level based on your current guard configuration and threat data.
* **Compliance reporting** — Generate reports mapped to OWASP LLM Top 10 categories and EU AI Act requirements directly from the Control Plane, using real threat event data from your production environment.
Because the Control Plane sits outside your AI workloads, it provides an authoritative and tamper-resistant record of security decisions — something no model-side safety mechanism can offer.
## Deployment Models
BeyondGuard supports two deployment modes. You can choose the mode that fits your current readiness level and switch between them at any time from the Control Plane.
**Inline (Enforcement) Mode** positions BeyondGuard as an active interceptor. Every interaction passes through BeyondGuard before reaching the model or the user. If a threat is detected and your policy says block, the interaction is stopped and a safe response is returned. This is the mode you use once you have calibrated your policies against real traffic and are ready for active protection.
**Observation Mode** positions BeyondGuard as a passive monitor. Interactions are inspected and logged, but BeyondGuard does not block or modify them. Threat events are recorded in the Control Plane for review, but your AI system's behavior is unchanged from the user's perspective. This is the mode you use to build a threat baseline, tune policies, and build organizational confidence before switching to enforcement.
Observation mode is the recommended starting point for most enterprise deployments. It lets you see exactly what BeyondGuard would have blocked — without affecting your users — so you can validate that your policies are correctly calibrated before activating enforcement. Most teams spend one to two weeks in observation mode before switching to inline enforcement on their first project.