Finding the first scenario
We record the business process, sources, users, roles, information security restrictions and success criteria.
Result: pilot scope and dependency map.Landing page
AI Security: Protecting LLMs, RAG, and Data in the Corporate Environment
AI security in a company involves controlling which data enters models, who access rights knowledge through AI assistants, and what remains in logs. RESTART performs AI system security audits and designs and implements a secure AI perimeter: protecting LLMs and RAG from data leaks and prompt injection, managing access rights, logs, and deploying models on-premise or in a private cloud.
AI System Security: What Are We Protecting Against
RESTART Group — security for corporate AI systems: audit, protection of LLMs and RAG, and control of AI-driven data leaks for companies with information security requirements.
AI system risks arise not in the model itself, but around it: in the data sent in queries, in the knowledge base index, in access rights, logs, and integrations. Below are typical threats and the measures RESTART builds into the architecture.
Data leakage through AI
PD, trade secret or code ends up in an external service, RAG index, prompt, response or journal without reason. Protection: Data classification, masking, DLP, private loop, retention policies and minimization.
Prompt injection
A document or user command causes the model to ignore rules or reveal data. Protection: Source filtering, system policies, sandbox for tools, red teaming and monitoring.
Bypassing rights in RAG
The assistant is showing an employee information from a document to which they do not have access. Protection: IAM/IDM rights inheritance, index filtering, source references, log verification.
Excessive authority of the agent
The AI agent is granted overly broad permissions and performs an action without coordination. Protection: Minimal permissions, action limits, human confirmation, logging — more about safe AI-agent automation.
Invalid answers
The model confidently answers without a source or mixes outdated documents. Protection: RAG with sources, quality gates, document versions, human verification in critical scenarios.
Attack surface
Public APIs, widgets, proxies, and external LLM services are exposed externally. Protection: Checking external points of the AI environment, pentest of AI/API perimeter.
Common vulnerabilities of LLM applications are described in OWASP Top 10 for LLM Applications, and AI system attack tactics are covered in MITRE ATLAS. In Russian practice, the AI perimeter is associated with Federal Law No. 152-FZ and Personal Data Processing Systems (ISPDn), requirements for Critical Information Infrastructure (CII) and State Information Systems (GIS), internal information security policies, and trade secrets.
AI Security Audit
If an AI pilot, RAG, chatbot, or external AI service is already in use, it's best to start with AI security audit — Secure AI Audit package. Within 10 business days, one agreed-upon AI pipeline is reviewed: what data is sent to the model and external services, who gains access to sources via AI, whether logs are complete, and where risks of data leaks, prompt injection, and vendor lock-in exist.
Risk register
Identified risks with significance assessment, priority, recommended action, and responsible role.
Architecture analysis
Where data goes beyond the boundary, where logs are missing, where roles are blurred.
Checklist
Requirements for data, access, and logs based on which the environment can be independently re-verified.
Road map
What to close before production operations, what can be postponed, and in what order.
Active testing — penetration testing of the AI/API perimeter — and verification of protection measures in cybersecurity labs are listed as separate items after the audit.
LLM and RAG Protection
LLM and RAG protection is built from multiple layers; no single component addresses all risks on its own.
Outline and Placement
Models and data — on-premise, in a private cloud, or hybrid. External LLM services — only after analyzing data, contractual terms, storage, and logging.
Access rights in RAG
User permissions are inherited from IAM/IDM, groups, and roles; the index is filtered, and the response references the source.
Masking and DLP
Sensitive fields are masked before entering the model and index — masking and depersonalization of data.
Prompt layer
Model, system instructions, response policies, tool restrictions, and filters—a separate controllable layer.
Logs and monitoring
Requests, responses, sources, errors, escalations, agent actions, and information security events are being logged.
Pre-launch check
Testing prompt injection, data leaks, access controls, agent permissions, logging, error handling behavior, and responses to controversial queries.
AI Safety and AI in Information Security
This page is about the security of AI systems themselves. The inverse task is to apply AI in information security: SOC and SGRC event analysis, preparation for audits under Federal Law No. 152-FZ, CIIP, and GIS. The Security & Compliance AI module is available for this purpose.
General information security practice — information security; the path from process assessment to AI deployment — implementation of AI in business.
Who is it suitable for?
CIO and CDTO
We need to move from chaotic AI pilots to a platform architecture that is clear to integration, operations and business.
CISO
It is important not to release data, documents, source code and personal information to uncontrolled external services.
Business owners
You need a quick effect in knowledge, documents, service desk, finance, procurement or development without loss of control.
Project office
It is necessary to formalize the roadmap, pilot KPIs, roles, budget outline and scaling rules.
Pain and management risk
The main problem with enterprise AI is not the lack of a model, but the lack of a managed architecture around data, access, quality and responsibility.
| Pain | What happens without preparation |
|---|---|
| AI pilots don't scale | Teams run individual scripts but fail to connect actual sources, roles, and auditing. |
| Data is getting out of control | Documents, trade secrets, personal data and code end up in services without a clear mode of use. |
| No trust in answers | The user receives text without sources, without checking the basis and without a clear process for improving quality. |
| CIO and CISO argue about risks | Architecture, information security, data and business impact are discussed separately, so the decision hangs. |
What RESTART Delivers
Designing a safe AI environment
We describe the target architecture: RAG, models, repositories, logs, integrations, access rights, dev/test/prod environments.
Result: HLD and platform requirements.Launching a controlled pilot
We connect a limited set of data, configure roles, prompts, quality gates and the feedback process.
Result: a working script with metrics.Preparing for production launch
We create a backlog, operational model, owners, risks, budget and replication plan.
Result: roadmap for 3-6 months.Artifacts at the output
AI opportunity map
Map of scenarios, priorities, owners and expected effects.
Secure AI architecture
Diagram of contours, sources, roles, logs, integrations and information security restrictions.
Pilot backlog
Pilot plan with quality criteria, data, metrics and acceptance.
Risk register
Risk register: personal data, trade secrets, hallucinations, access, operation, vendor lock-in.
First stage deadline
The optimal start is AI-discovery in 2 weeks. This is enough to separate real scenarios from beautiful ideas, evaluate the data and make management decisions about the pilot.
Team composition
| Role | Area of responsibility |
|---|---|
| AI architect | AI platform, RAG, models, orchestration, quality criteria. |
| Data architect | Sources, data quality, vectorization, DWH/BI/ERP integration. |
| Security architect | PD, access rights, logs, contours, threats and information security requirements. |
| Business analyst / PM | Process, stakeholders, backlog, acceptance, roadmap and communications. |
Standard plan
1
Context and goals
Interviews with CIO, CISO and process owners, recording goals and limitations.
2
Data and Risks
Assessment of sources, access, data quality, personal data, trade secrets and information security risks.
3
Pilot architecture
Design of RAG/AI loop, integrations, roles, logs and quality criteria.
4
Roadmap
The pilot's plan, team, timelines, artifacts, budget guidelines and the next stage.
Relevant cases and experience
Public cases do not reveal confidential details of customers, but show RESTART’s experience in related enterprise landscapes.
Boundary with action automation
Here we're talking about the perimeter: sources, models, roles, logs, deployment, and operation. As soon as the assistant stops responding with text and starts performing actions within systems—creating tickets, modifying records, calling external APIs—a separate discipline kicks in: agent permissions, human confirmations, and rollback capability.
This part is described on the page safe automation of processes by AI agentsUsually, the order is as follows: first, the environment is built, then the agent is granted rights to perform actions.
Frequently asked questions
What is AI system safety?
These are measures that prevent corporate AI from becoming a channel of leakage or attack: control of data entering the model and index, inheritance of access rights, request and response logs, protection from prompt injection, limiting the powers of AI agents and human verification in critical scenarios.
How to prevent data leaks through AI?
Classify data, mask sensitive fields, use DLP, keep models and indexes within a private perimeter, define policies for storing queries and responses, and transmit only the minimally necessary information to the model.
How to protect an LLM from prompt injection?
Filter sources, enforce system policies, sandbox agent tools, perform red teaming, and monitor logs. Before running prompt injection, it is verified by tests.
How to protect RAG from leaks?
Requires source classification, access rights inheritance, index filtering, sensitive data masking, response control, source references, and log inspection.
What does an AI security audit check?
Secure AI Audit checks one AI system within 10 business days: data, access rights, architecture, logs, integrations, and risks—including personal data and trade secrets in queries, prompt injection, and vendor lock-in. The deliverable is a risk register and a remediation roadmap.
Is it possible to start without a platform?
Yes. The first stage is needed just to choose an architecture: a separate pilot, a platform core or a hybrid path.
Is it possible to work without transferring data to an external AI service?
Yes. The architecture can be on-prem, private cloud or hybrid, with source limitation, logging and access control.
What is considered a pilot's success?
Not the number of requests, but the quality of answers from sources, reduction of manual labor, clear risks and readiness for industrial use.
The entire path - from process inspection to operation - is described on the page enterprise AI for business.
Email us
Write to us at info@restart.re or click the button below. For the first letter, it is sufficient to briefly describe the outline, the role of the organization, current systems, limitations and the desired outcome.
Email usRequest format
We agree on the volume, timing and output artifact - and launch it in a short time.
