Core concepts, in plain terms

An AI model trained on large volumes of text that can generate natural-language responses. It is often the starting point of enterprise AI architecture, but on its own it does not represent enterprise capacity.

A category of AI systems that produce new text, images, audio, or code; unlike classic prediction/classification models, they generate probabilistic, creative output.

A large-scale model trained on broad, general-purpose data and adaptable to many tasks; most enterprise applications either use these models directly or customize them.

An AI component that can interpret a goal, break it into sub-steps, query enterprise data sources, call tools, and complete transactions under defined conditions. → Who It's For / What Is EnaGuard

An architecture in which multiple agents share tasks and information to work together; it creates more complex coordination and authorization requirements than a single-agent model.

An AI system that processes text, image, audio, video, and sensor data within the same workflow; each data type carries different storage, privacy, and security requirements.

Unlike large general-purpose models, a smaller model trained or adapted to produce faster, more economical, and more consistent results in a specific task or domain.

The process of further training a foundation model to adapt it to a specific task or an organization's own data set.

An optimization technique that reduces a model's numeric precision so it can run with fewer resources and lower latency.

An AI model that runs close to where data is generated (field equipment, a production line, a vehicle, a store); it arises from the need for low latency and offline operation.

Artificially generated data that carries statistical properties similar to real data; it can be used for model training or testing where sensitive-data constraints apply.

An architectural approach that lets a model retrieve and use relevant information from enterprise sources when generating a response. → What Is EnaGuard

A database that turns text into numeric vectors and performs semantic similarity search across them; a core component of RAG systems.

Re-ordering the candidates returned by an initial search using a stronger model; it directly affects RAG answer quality.

The process of splitting a document into smaller, meaningful pieces before indexing; chunking quality affects search accuracy.

A layer between applications and model services that centralizes access, routing, policy enforcement, cost control, and logging. → What Is EnaGuard

Directing a request to the most suitable model based on use case, data class, latency target, or cost limit.

An integration protocol that lets models and agents connect to data sources and tools through a common method.

The discipline of managing models, prompts, data sets, and configurations in a traceable way across the development, testing, release, and production lifecycle.

The principle that an agent or user should hold only the data and transaction rights required for its task.

The digital identity an agent or system component uses to access enterprise resources; each agent should ideally hold a separate, traceable service account.

Access credentials valid for a limited time and renewed automatically, instead of long-lived passwords or keys; this limits the impact of a leak.

A design principle requiring human approval for actions that cross a defined risk threshold.

A mechanism that quickly stops an agent and revokes its access when unexpected behavior, an authorization breach, or a faulty decision chain is detected.

A resilience pattern that temporarily cuts calls to a failing dependency to prevent the fault from cascading through the system.

A design that lets a service continue via a secondary model, a limited function, or a manual process when the primary model or provider is unavailable.

Controlled re-attempting of a request after a transient error; because uncontrolled retries can amplify a problem, it requires a centralized, bounded policy.

An attempt to hijack model behavior via instructions embedded in user input or processed content such as a document, web page, email, or tool output.

A protection layer that checks a model's inputs and outputs against acceptable behavior boundaries.

Independent adversarial testing that uncovers weaknesses in a system before a real attacker does.

Deliberately manipulating training data, fine-tuning data, or RAG sources to corrupt model behavior.

The field of attack and defense that uses inputs specifically crafted to mislead a model's decision mechanism.

Copying a model's behavior from its outputs (extraction), or inferring information about its training data (inversion); both carry intellectual-property and privacy risk.

Employees using third-party AI tools without the organization's knowledge or approval; it can make corporate data flows invisible to the organization.

Data that can directly or indirectly identify a person (such as a name, national ID, or email); it requires special protection under KVKK and GDPR.

A capability that the team simply states exists; supporting evidence is still limited.

The architecture, policy, procedure, or standard has been defined; it may not yet be implemented in production.

The relevant technology or control has been deployed and configured.

The capability has been validated under normal, load, failure, and attack scenarios.

A capability level backed by production logs, metrics, test records, and incident data, monitored and improved on an ongoing basis.

Not merely that an AI system runs, but that it can be operated safely, scalably, and sustainably under real production conditions.

The verification of an organization's declared AI architecture capabilities, with evidence, by an independent party free of conflicts of interest. → What Is EnaGuard

A control is declared to exist, but its scope, ownership, production linkage, or test status is unclear, so the real state stays invisible.

An evidence-based view of maturity, evidence confidence, scope coverage, residual risk, QA outcome, and claim boundary; not a single score.

When the person submitting evidence and the person who would review it are the same, a separate reviewer must sign off before that evidence counts; without it, confidence in the result is reduced.

An international standard for an AI Management System (AIMS); it aims to help organizations manage AI risks systematically.

An international standard on AI risk management.

A voluntary AI risk-management framework from the US National Institute of Standards and Technology, built around the Govern / Map / Measure / Manage functions.

An open-source, globally contributed security research initiative focused on the security risks of LLM- and agent-based systems.

A continuously updated knowledge base compiling real-world attack tactics and techniques against AI systems.

A cloud and security control framework for AI systems developed by the Cloud Security Alliance.

The European Union's risk-based AI regulation; it imposes heavier obligations on high-risk systems.

Türkiye's and the European Union's core personal-data-protection regulations; they directly shape how AI systems may process data.

The country or region where data is processed and stored; an architectural parameter that directly shapes regulatory, latency, and security decisions.

A deployment model in which AI workloads run primarily on cloud platforms, under relatively flexible regulatory/data constraints.

A deployment model in which some workloads run in the cloud and others on internal infrastructure, under restricted cloud usage.

A deployment model fully isolated from external networks, typically chosen for high-confidentiality or critical-infrastructure requirements.

A deployment model in which AI infrastructure is hosted in the organization's own data center.