> ## 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. # Frequently Asked Questions: BeyondGuard AI Security > Find answers to the most common questions about BeyondGuard's guards, deployment options, compliance support, and enterprise AI security approach. The questions below cover what new and prospective BeyondGuard customers ask most often — from teams evaluating AI security for the first time to security architects designing enterprise-scale deployments. If you do not find the answer you need here, reach out to the BeyondGuard team directly through the Control Plane or contact your account representative. No. BeyondGuard adds the AI-specific security layer that traditional tools — firewalls, DLP platforms, CASB solutions — were not designed to provide. Conventional security tooling operates on network traffic, file transfers, and known malware signatures. It has no visibility into the semantic content of LLM interactions, cannot detect prompt injection, and cannot evaluate whether an AI agent's actions are within its intended scope. BeyondGuard complements your existing security stack rather than replacing it. It sits specifically at the AI layer, covering the threat surface that your current tools cannot reach. Model-level guardrails are built into the model itself — they are part of the same system that adversarial prompts are attempting to manipulate. A sufficiently sophisticated jailbreak or prompt injection attack can bypass these guardrails by operating on the same input channel that the guardrails use to make decisions. This is not a hypothetical: new jailbreak techniques that circumvent model-level safety training are discovered regularly. BeyondGuard is structurally independent. It operates outside the model as a separate control plane, evaluating interactions before they reach the model and after responses leave it. Its detection and enforcement decisions cannot be overridden by adversarial content in the user prompt — because that content never reaches the BeyondGuard evaluation layer in a form that can manipulate it. This structural independence is the foundation of BeyondGuard's security model. No. You can begin your BeyondGuard engagement with a [Red Teaming Assessment](/guides/red-teaming), which requires no runtime deployment and no changes to your existing AI infrastructure. A Red Teaming Assessment gives you a scored, evidence-based view of your current exposure across the full AI threat model — without touching production systems. When you are ready to deploy runtime controls, getting started requires adding the BeyondGuard SDK or API integration to your application and enabling your first guard in Observation Mode. Most customers complete initial Prompt Guard deployment in less than a day. Start with Prompt Guard in Observation Mode. Prompt Guard has the broadest coverage of any single guard — it addresses prompt injection, jailbreaks, system prompt extraction, and out-of-scope requests, which together account for the majority of AI application attacks — and it has the lowest deployment complexity of any BeyondGuard guard. Deploying Prompt Guard in Observation Mode first gives you a threat baseline for your specific application before you commit to any enforcement configuration. Once you have that baseline and have tuned your policies against real traffic data, you can switch to Enforcement Mode and expand to Agent Guard, MCP Guard, Data Guard, and Output Guard in sequence. Yes. Prompt Guard includes multi-modal attack detection for AI models that accept images, audio, or documents as input alongside text. Multi-modal inputs introduce attack vectors that do not exist in text-only systems — adversarial images containing embedded instructions, documents with hidden text layers, and audio inputs designed to trigger specific model behaviors. BeyondGuard's Prompt Guard evaluates these input types for injection and manipulation patterns as part of standard guard operation. BeyondGuard evaluates AI interactions through a structurally independent control plane. In standard inspection modes, your interaction data does not need to leave your environment — BeyondGuard can be deployed in configurations that keep all data within your infrastructure boundary. Specific deployment architectures vary depending on your environment, your regulatory requirements, and your data residency obligations. Contact BeyondGuard to discuss the deployment architecture that is right for your organization, including options for on-premise and private cloud deployments. BeyondGuard is designed for enterprise deployments in demanding, regulated sectors — including banking and other regulated financial services — where data residency, network isolation, and infrastructure control requirements are strict. BeyondGuard supports deployment architectures tailored to these constraints. Contact the BeyondGuard team to discuss the specific requirements of your environment. Whether you need an on-premise deployment, a private cloud configuration, or a hybrid architecture, the BeyondGuard team can work through the options with you. A [Red Teaming Assessment](/guides/red-teaming) gives you a quantified baseline: a scored vulnerability report that identifies your current exposure by severity and category before any guards are deployed. This baseline serves as the starting point for measuring improvement. After deploying BeyondGuard guards, the Threat Events dashboard shows the volume and severity of threats intercepted in real time — a direct, ongoing measure of the value being delivered. You can also request follow-up Red Teaming Assessments at intervals to measure how your risk score has changed as your guard coverage and policy maturity have developed. For organizations at L4 or L5 maturity, BeyondGuard's compliance dashboards generate structured reports that quantify threat coverage, enforcement activity, and policy effectiveness over time — providing the data that security and executive leadership need to evaluate the program's impact. Yes. BeyondGuard provides the audit trails, enforcement evidence, and policy documentation that support compliance with EU AI Act obligations for high-risk AI systems. Specifically, BeyondGuard generates: * Comprehensive audit logs of all AI interactions and policy decisions * Evidence of active cybersecurity controls as required for high-risk AI system classification * Threat event records demonstrating that anomalous behaviors were detected and addressed * Policy documentation showing that defined acceptable use rules were enforced See the [Compliance guide](/guides/compliance) for a detailed mapping of BeyondGuard controls to EU AI Act obligations and OWASP LLM Top 10 categories. Most customers complete initial Prompt Guard deployment in less than a day. The process involves creating a project in the BeyondGuard Control Plane, enabling Prompt Guard in Observation Mode, and integrating the BeyondGuard SDK into your application — a straightforward API integration that typically takes a few hours for an engineer familiar with the codebase. See the [Deploy BeyondGuard Across Your AI Stack](/guides/deploying-guards) guide for the complete step-by-step process, including the SDK integration code and guidance on reviewing your first threat events.