Glossary · Category

Deployment & security

110 plain-English definitions.

A/B testing AI models: Comparing two model versions or providers against each other on live traffic to see which performs better
agentic AI governance: Policies and technical controls specifically designed to manage the risks of autonomous, multi-step AI agents
AI agent orchestration: Coordinating multiple AI agents and tools to complete complex, multi-step business workflows
AI agent security: Controls and safeguards that prevent autonomous AI agents from taking unintended or harmful actions
AI API access control: Restricting who and what systems can call an organization's AI endpoints, and with what permissions
AI content filtering enterprise: Automated screening of AI-generated or AI-processed content to block sensitive or inappropriate material
AI credential rotation: Regularly changing API keys and secrets used to access AI systems to limit exposure from leaks
AI data anonymization: Removing or obfuscating identifying details from data before it is used to train or query an AI model
AI data encryption at rest: Encrypting stored data used or produced by an AI system so it is unreadable if the storage is compromised
AI data encryption in transit: Encrypting data as it moves between a client and an AI system to prevent interception
AI data flow mapping: Documenting exactly how data moves into, through, and out of an AI system for security and compliance review
AI data residency requirements: Rules dictating which physical location an AI system's data must be stored or processed in
AI DevSecOps: Integrating security practices directly into the development and deployment pipeline for AI systems
AI edge deployment: Running AI inference directly on local devices rather than sending data to a remote server
AI endpoint security: Protecting the devices and interfaces employees use to access AI tools from compromise or misuse
AI failover strategy: A backup plan that reroutes AI requests to an alternate model or provider if the primary one goes down
AI guardrails: Technical controls that constrain an AI system's behavior to prevent harmful, off-topic, or noncompliant outputs
AI infrastructure as code: Managing AI deployment configuration through version-controlled code rather than manual setup
AI infrastructure resilience: The ability of an AI system to keep functioning correctly despite failures, spikes in load, or outages
AI infrastructure scaling: Adding or reducing compute capacity to match the changing demand of AI workloads
AI model access governance: Controlling and auditing who within an organization can use, configure, or query specific AI models
AI model access logging: Recording who accessed which AI model and when, for security and compliance purposes
AI model API key rotation policy: A defined schedule and process for regularly changing the credentials used to access AI provider APIs
AI model backup and restore: Procedures for preserving and recovering model configurations and fine-tuned weights
AI model context window management: Techniques for efficiently managing how much text and history is fed into a model within its context limit
AI model endpoint protection: Security controls specifically guarding the API endpoints through which a model is accessed
AI model isolation: Architecting AI systems so one tenant's or application's model instance cannot access another's data
AI model self-hosting vs API tradeoff: The decision between running a model on owned infrastructure versus paying for hosted API access
AI model serving cost vs latency tradeoff: The balance between paying more for faster model responses and saving money with slower, cheaper serving
AI model uptime guarantee: A vendor's contractual commitment to how reliably its hosted model will remain available
AI model versioning: Tracking and managing different iterations of a model so deployments can be rolled back or compared
AI network segmentation: Isolating AI systems on separate network segments to limit the blast radius of a security breach
AI observability: Monitoring and tracing tools that give visibility into how AI systems are performing and behaving in production
AI penetration testing: Simulated attacks against an AI system used to uncover security vulnerabilities before real attackers do
AI request logging and audit: Recording every AI system request and response for later review, debugging, or compliance purposes
AI request routing: Directing each AI request to the specific model best suited to handle it, based on complexity, cost, or capability
AI request throttling: Deliberately slowing or capping the rate of AI requests to control cost or protect system stability
AI request tracing: Following a single request through every system it touches to debug issues or verify compliance
AI supply chain security: Verifying that the models, libraries, and datasets an AI system depends on have not been tampered with
AI system audit logging: Recording detailed, immutable logs of every action an AI system takes for later security or compliance review
AI system hardening: Configuring an AI deployment to minimize its attack surface and reduce vulnerability to exploitation
AI system uptime monitoring: Continuously tracking whether an AI service is available and responding as expected
AI tool permissioning: Defining exactly which systems and actions an AI agent is allowed to access or perform
AI uptime SLA: A vendor's contractual commitment to a minimum percentage of time its AI service will be available
AI usage anomaly detection: Automated identification of unusual patterns in how an AI system is being used, which may indicate misuse
air-gapped AI: An AI system deployed on infrastructure with no connection to the public internet, used for maximum security
API gateway for AI: A managed entry point that handles authentication, rate limiting, and routing for all AI API traffic
bring your own key: An arrangement where a customer supplies their own encryption key so a vendor cannot access their data
bring your own model: An enterprise arrangement allowing a company to plug in its own preferred AI model into a vendor's platform
BYOC AI: Bring your own cloud, an arrangement allowing a customer to run a vendor's AI software inside their own cloud account
canary deployment AI model: Gradually rolling out a new model version to a small subset of traffic before a full release
confidential computing AI: Hardware-based techniques that keep data encrypted even while an AI model is actively processing it
data loss prevention AI: Security controls that stop sensitive data from being exposed or exfiltrated through AI tool usage
disaster recovery AI systems: Plans and procedures for restoring AI services and data after a major outage or failure
disaster recovery for AI models: Plans for restoring AI model availability and data after a catastrophic infrastructure failure
enterprise AI change management for deployments: Formal processes for reviewing and approving changes to production AI systems before they go live
enterprise AI data pipeline security: Protecting the flow of data feeding into and out of AI systems from unauthorized access or tampering
enterprise AI deployment architecture: The overall technical design of how AI systems are integrated into a company's infrastructure
enterprise AI deployment checklist: A structured list of steps and requirements to complete before putting an AI system into production
enterprise AI DR and business continuity: Planning to keep critical AI-dependent business processes running through an outage or disaster
enterprise AI key management: Securely generating, storing, and rotating the credentials used to authenticate with AI providers
enterprise AI security: The practices and controls that protect an organization's AI systems, models, and data from compromise
enterprise AI service mesh: Infrastructure that manages secure, observable communication between the many services that make up an AI application
enterprise AI zero trust deployment: Deploying AI systems under a security model that verifies every access request rather than assuming trust
enterprise API integration AI: Connecting an AI system to a company's existing business applications through their APIs
enterprise identity and access management AI: Integrating AI system access with a company's central identity provider for consistent authentication and authorization
enterprise LLM deployment patterns: Common architectural approaches companies use to put large language models into production
enterprise LLM gateway architecture: The technical design of a middleware layer that centralizes routing, security, and monitoring for all AI traffic
enterprise LLM integration: The work of connecting a large language model into a company's existing applications and workflows
enterprise self-hosting AI: The practice of running AI models within a company's own controlled environment instead of using external APIs
fallback model strategy: A plan to automatically switch to a backup model if the primary model fails or is unavailable
federated learning enterprise: A technique for training AI models across decentralized data sources without moving the raw data to a central location
hybrid AI deployment: An architecture that splits AI workloads between on-premise infrastructure and cloud services
intelligent model routing: Automatically selecting the most appropriate and cost-effective model for each individual request
LLM abuse prevention: Techniques for detecting and blocking malicious or policy-violating use of an AI system
LLM error handling: Strategies for gracefully managing failures, timeouts, or malformed outputs from a language model
LLM latency optimization: Techniques used to reduce the time it takes a language model to return a response
LLM load balancing: Distributing AI requests across multiple models or provider endpoints to optimize speed, cost, and reliability
LLM monitoring: Tracking the latency, cost, quality, and error rate of large language model requests in production
LLM output validation: Automatically checking an AI-generated response for correctness or policy compliance before it's used downstream
LLM proxy: An intermediary service that forwards requests to language models while adding logging, caching, or access control
LLM sandboxing: Running AI-generated code or actions in an isolated environment to contain any unintended or malicious behavior
LLM security best practices: Recommended controls for safely deploying large language models in production, covering input validation, access control, and monitoring
MCP server enterprise deployment: Running a Model Context Protocol server within a company's infrastructure to safely expose internal systems to AI agents
model context isolation: Ensuring one user's or tenant's conversation context cannot leak into another's session
model context protocol enterprise: An open standard, introduced by Anthropic, that lets AI assistants securely connect to a company's internal tools and data sources
model output filtering: Automatically screening AI-generated content for policy violations, sensitive data, or unsafe content before it reaches users
model quantization enterprise: Compressing a model to run faster and cheaper with minimal loss in accuracy
model rollback: Reverting a production AI system to a previous, known-good version of a model after a problem is discovered
model serving infrastructure: The technical stack used to host and run a trained model so it can respond to real-time requests
model weight security: Protecting a trained model's parameters from theft, since they represent significant intellectual property
multi-provider AI redundancy: Using more than one AI vendor so a failure or outage at one provider does not halt operations
multi-tenant AI security: Ensuring that a shared AI service keeps different customers' data and requests fully isolated from each other
on-prem AI deployment: Installing and operating AI infrastructure within a company's own data centers
on-premise LLM: A large language model hosted and run on a company's own servers rather than a third-party cloud API
PII redaction AI: Automatically detecting and removing personally identifiable information from text before or after it is processed by an AI model
private AI deployment: Running AI models within an organization's own infrastructure so data never leaves its control
private cloud AI hosting: Running AI infrastructure in a dedicated, single-tenant cloud environment rather than shared public cloud
prompt injection attack: A security exploit where crafted input manipulates an AI model into ignoring its original instructions
prompt injection defense: Techniques used to detect and block attempts to manipulate an AI model through malicious input
rate limiting AI API: Restricting how many requests a client can send to an AI service in a given time window
role-based access control AI: Limiting AI system permissions and data visibility based on a user's assigned role within the organization
secure AI API keys management: Practices for generating, rotating, and protecting the credentials used to authenticate with AI services
secure AI sandbox: An isolated environment used to safely test AI systems before they touch production data or systems
secure enclave AI processing: Using isolated, hardware-protected computing environments to process sensitive data with AI
secure prompt handling: Practices for ensuring prompts containing sensitive information are transmitted and stored securely
self-hosted LLM: An open-weight language model that a company downloads and runs on its own infrastructure
single sign-on AI tools: Allowing employees to access AI applications using their existing corporate login credentials
zero data retention AI: An AI provider's commitment not to store or reuse a customer's input and output data after processing
zero trust AI architecture: A security model that requires every AI system interaction to be verified, with no implicit trust granted