Integrating large language models (LLMs) and autonomous AI agents into enterprise infrastructure introduces new security challenges. According to Microsoft, 49% of business chats in Copilot are involved in complex cognitive tasks. However, traditional security tools like Web Application Firewalls (WAFs) are ineffective against semantic prompt manipulations, as the boundary between code and data is blurred within LLM architecture.
Critical threats to enterprise AI
According to the OWASP 2025 classification, the primary threats to AI systems include:
- Prompt Injection (LLM01:2025): manipulating input text to bypass system instructions.
- Sensitive Information Disclosure (LLM02:2025): the leakage of confidential data through model responses, particularly in Retrieval-Augmented Generation (RAG) architectures.
- Indirect Prompt Injection: embedding malicious instructions through external sources read by the AI agent.
According to Cisco, only 13% of organizations are prepared to counter such threats at the architectural level.
The effect on the market
For enterprises, the vulnerability of LLMs to semantic attacks means that traditional cybersecurity perimeters are no longer sufficient. Organizations risk massive data leaks, unauthorized access to internal databases, and compliance violations if they deploy AI agents without specialized safeguards. With only 13% of organizations prepared, the industry faces an urgent need to transition from network-level security to semantic and data-level defense architectures.
What to do: Architectural defense under NIST and OWASP standards
To minimize risks, experts recommend implementing the NIST AI RMF 1.0 framework, which outlines four functions: Govern, Map, Measure, and Manage.
Practical infrastructure defense is based on the following steps:
- Data isolation at the DBMS level: applying Row-Level Security (RLS) and Role-Based Access Control (RBAC) before information enters the model's context.
- Separation of instructions and data: using structured API templates to distinguish between system and user roles.
- Validation and filtering: automatically masking personally identifiable information (PII) in the model's output using Guardrails systems.
- AI Red Teaming: regularly testing system resilience using the MITRE ATLAS methodology.
Practical implementation: Ukrainian experience
An example of secure integration is the solution from Softengi, which implements AI systems following Security by Design principles on the UnityBase low-code platform by InBase. Thanks to built-in RLS support and database-level auditing, the RAG system is physically unable to access confidential documents for which the user lacks permissions, even in the event of a successful Prompt Injection attack.
Prepared by a Software Ukraine member. Original publication.