Glossary

Definitions and practical examples for the concepts behind AI agents, identity, and governance.

ABAC vs. RBAC: Role-based access control, or RBAC, assigns permissions through roles.
Agent audit trails vs. LLM observability: An agent audit trail is a set of retained records used to reconstruct what an agent did and which decisions allowed or stopped it.
Agent authorization: Agent authorization determines whether an identified agent may perform a requested operation on a resource under the applicable conditions.
Agent incident response: Agent incident response is the coordinated work of detecting, investigating, containing, and recovering from harmful or unauthorized activity involving an agent.
Agent lifecycle management: Agent lifecycle management is the work of introducing, changing, suspending, and retiring an agent's operational identity and its associated access.
Agent registry vs. agent inventory: An agent inventory is a list of agents or agent workloads believed to exist.
Agentic operations: AgentOps, LLMOps, and AIOps: Agentic operations is the work of running agent workloads as services with named owners: introducing them, controlling their access, observing their activity, changing them, and retiring them.
AI agent security: AI agent security protects the software, identities, data, and connected systems involved when an agent acts.
AI gateway vs. LLM gateway vs. MCP gateway: An AI gateway sits between an AI application and the services it calls.
AI governance vs. agent governance: AI governance is the organizational work of deciding how AI may be developed, acquired, used, and reviewed.
AI TRiSM: AI TRiSM stands for AI trust, risk, and security management.
Authentication vs. authorization: OAuth and OIDC: Authentication establishes the identity associated with a request.
Content filtering vs. schema validation: Content filtering inspects content and may transform, withhold, or block it according to configured criteria.
Cost allocation: AI cost allocation assigns direct and shared AI-related costs to teams, products, customers, or workloads using an explicit method.
Credential isolation: Credential isolation keeps a service credential out of the calling application's runtime.
Data redaction vs. masking and omission: Data redaction removes or obscures information before it is disclosed.
Default-deny vs. fail-closed: Default-deny means a request is refused when no applicable permission allows it.
Delegated authorization: Delegated authorization lets one actor perform specified operations on behalf of another.
Guardrails: AI guardrails are controls intended to keep an AI application's inputs, outputs, or actions within defined constraints.
Least privilege for AI agents: Least privilege limits an actor's authority to what is needed for its assigned work.
LLM failover vs. fallback: LLM failover moves a request to another serving destination after a defined failure.
LLM routing and model routing policy: LLM routing, also called model routing, selects a model or serving endpoint for a request.
MCP (Model Context Protocol): The Model Context Protocol (MCP) defines how AI applications connect to external context and capabilities.
MCP security: Model Context Protocol (MCP) security protects the connections, credentials, capabilities, and data involved in an MCP deployment.
Multi-provider AI: Multi-provider AI uses AI services from more than one provider within an application or operating environment.
Non-human identity vs. workload identity: A non-human identity represents an actor other than a person, such as software or a device.
Policy as code: Policy as code expresses requirements as rules that software can evaluate.
Policy testing vs. policy simulation: Policy testing evaluates rules against selected inputs and expected outcomes.
Prompt injection: direct and indirect: Prompt injection occurs when content supplied to a language-model application redirects behavior away from the application's intended instructions or task.
Rate limits vs. quotas and budgets: Rate limiting constrains how quickly requests or units of work may be consumed.
Runtime policy enforcement: Runtime policy enforcement applies a policy decision while an operation is being processed, through a component that can allow, modify, or stop the operation.
Spend governance: AI spend governance defines how an organization approves, controls, and reviews spending on AI.
Tool calling and function calling: Tool calling is the mechanism through which an AI application requests an external operation, such as looking up a record or creating a calendar event.
Tool poisoning: Tool poisoning places malicious or misleading instructions in material describing a tool, including descriptions and parameter metadata, so an AI application selects or uses capabilities in an unintended way.
Unit economics: AI unit economics relates the cost of AI-enabled work to a defined unit of output or value.
Zero trust for AI agents: Zero trust is an approach to security that does not grant implicit trust solely because an actor or resource is inside a network or owned by the organization.