As AI integration into corporate systems accelerates, cybersecurity leaders must shift their approach to protection. Treating AI models as closed "black boxes" and deploying isolated tools creates blind spots and uncontrolled data leaks. Today, CISO and CTO roles require a shift in focus: AI components must be treated as standard infrastructure elements requiring strict architectural control and risk management according to proven standards.
Moving from AI security to architectural security
Organizations often seek specific solutions to protect algorithms while ignoring fundamental network design principles. According to the ENISA Threat Landscape 2025 report, traditional attack vectors—such as phishing and supply chain compromise—remain primary threats despite AI evolution. AI security in the corporate sector is achieved not through separate "AI-specific" add-ons, but by integrating models into a classic security architecture.
AI agents and LLM services often receive excessive access rights to corporate databases and APIs without proper authentication for each request. This creates new compromise vectors. Statistically, about 53.7% of incidents involving data leaks through automated systems occur due to a lack of access control at the level of algorithm interaction with internal sources. An attacker only needs to use prompt injection to bypass model logic and access sensitive information if the architecture does not restrict its permissions.
Applying the NIST AI RMF 1.0 framework to cybersecurity
To address these challenges, the NIST AI RMF 1.0 (Artificial Intelligence Risk Management Framework) is used. This document structures AI risk management around four key functions: Govern, Map, Measure, and Manage. NIST emphasizes that when implementing AI in critical infrastructure, organizations must evaluate context, reliability, security, and potential harm, rather than just the mathematical accuracy of the algorithm.
This approach aligns with the updated NIST Cybersecurity Framework (CSF) 2.0, which introduces the "Govern" function for deep integration of cyber risk management into the overall corporate governance structure. AI integration efforts should begin with the Govern and Map functions to identify threats early in the lifecycle. Currently, only 27.7% of organizations have a complete data flow map for their AI systems, leaving them vulnerable to unauthorized access.
Architectural control: Implementing Zero Trust and RBAC for AI agents
To protect infrastructure from leaks via models, strict application of proven access control practices is necessary:
- Role-Based Access Control (RBAC): Applying RBAC to AI-based analytical tools ensures that an automated agent has access only to data permitted for the authorized personnel on whose behalf it acts.
- Zero Trust: Using Zero Trust principles means mandatory verification of every request from an AI agent to internal databases and APIs. Automatic trust in internal components is unacceptable.
- Row-Level Security (RLS): Implementing access checks at the individual database record level (RLS) makes it impossible to extract sensitive rows, even if an attacker manages to manipulate the model prompt.
Audit and traceability: Why logging model decisions is critical
It is impossible to manage a system whose decisions cannot be tracked. Implementing detailed audit trails for AI-generated decisions is a fundamental requirement for accountability. Logging must cover input prompts, model responses, confidence levels, and request context. Assessing the resilience of AI components to manipulation and systematically checking access logs directly fulfills the Measure function within the NIST AI RMF framework.
ISO/IEC 42001 and threat readiness: Assessing corporate system maturity
According to the Cisco Cybersecurity Readiness Index 2025, organizational readiness assessment now includes "AI Fortification" as a key security pillar alongside network identification and resilience. While compliance with standards does not guarantee absolute immunity from cyberattacks, it demonstrates a mature risk management process.
For the development of secure AI agents and seamless integration of AI solutions into corporate infrastructure, the expertise of Softengi (part of the Intecracy Group technology alliance) is utilized. Softengi is certified to the ISO/IEC 42001:2023 AI management standard, allowing for the construction of custom AI systems and cloud integration for the enterprise sector, taking architectural security requirements into account. Combined with platform mechanisms such as UnityBase (which provides a unified domain model, RBAC/RLS, and end-to-end auditing), this ensures controlled deployment of innovations without compromising corporate data.
| Function of NIST AI RMF | Classic control tool | Practical implementation for AI |
|---|---|---|
| Govern | Corporate policies and data classification | Defining allowed data types for training and LLM queries |
| Map | Asset inventory and threat modeling | Mapping all AI model and API data entry/exit points (Data Flow) |
| Measure | Penetration testing and auditing | Assessing AI component resilience to manipulation and checking access logs |
| Manage | Access control (RBAC, Zero Trust) | Implementing row-level access checks (RLS) for every AI agent request |
AI security requires moving away from searching for a unique "silver bullet." Organizations must focus on aligning their technological processes with NIST AI RMF and integrating AI components under the umbrella of Zero Trust architecture and strict logging.
FAQ
How to integrate NIST AI RMF 1.0 into an existing enterprise cybersecurity system?
Integration begins by mapping the four functions (Govern, Map, Measure, Manage) onto existing risk management processes. AI models are treated as standard assets: a data flow map is created for them, information classification policies are applied, penetration testing is conducted, and Zero Trust principles are implemented.
How to protect sensitive corporate data from leaks via public and private LLMs?
Protection is implemented at the architectural level by applying Zero Trust principles, Role-Based Access Control (RBAC), and Row-Level Security (RLS). Every request from an AI model to a corporate database must be authenticated and verified just like a request from a regular user.
Which ISO/IEC 42001 standard requirements are key for a CISO when implementing AI?
The key is establishing a formalized Artificial Intelligence Management System (AIMS). It requires defining security policies for AI, assessing the context and potential harm of model usage, ensuring decision traceability through audit trails, and managing the algorithm lifecycle.