GLOSSARY

Agentic Security & Governance Glossary

AI agents now call tools and hand work to other agents, often under credentials that were designed for people or ordinary applications. Security and identity teams need a shared vocabulary for what changes when software can act on its own initiative.

This glossary defines the terms we use at Klyn. Each entry gives a plain definition, followed by the Klyn perspective: how we think about the problem and where Klyn applies. Where a capability depends on a supported integration or a protected enforcement point, we say so.

Agents and the agentic estate

AI agent

Software that uses a language model to work toward a goal by taking actions through tools and APIs. An agent differs from a chatbot because its output can change something: a record, a configuration, a payment or another system's state.

Klyn perspective: We treat every agent as an actor that needs an identity, an accountable owner and limits on what it may do. The useful question is what an agent can reach and change.

Agentic AI

The class of AI systems that pursue goals across multiple steps with some autonomy, using tools and other agents along the way. Earlier generative AI mostly produced text for a person to review; agentic systems act on what they produce.

Klyn perspective: Autonomy moves the security boundary. When a model can choose its next action, controls have to sit where the action happens, outside the model's reasoning.

Agentic estate

The full population of AI agents in an organization, together with the tools, data, credentials and workflows they use. It spans agents running on employee machines, in cloud services, inside SaaS platforms and in code repositories.

Klyn perspective: Klyn Connect brings endpoint, sensor and supported application, cloud and identity signals into one connected view of the estate, so teams can see where agents run and how the work behind a business task moves between them.

Agent sprawl

The uncontrolled growth of AI agents across business systems without clear ownership, inventory or oversight. Sprawl is usually a by-product of how easy agents are to build: a SaaS admin panel or a laptop is enough to start one.

Klyn perspective: Sprawl is an ownership problem before it becomes a security incident. Klyn discovers agents across connected sources and maps each one to its owners and relationships, so nothing important runs without someone answerable for it.

Shadow AI

AI tools and agents that people use or build without the knowledge or approval of security and IT teams. Shadow AI often holds real access, such as OAuth grants or MCP connections, that nobody has reviewed.

Klyn perspective: Agentic discovery surfaces new agents in the Klyn inventory as they appear in connected sources, along with what they can already reach. Finding them early turns an unknown risk into a review task.

Identity and ownership

Agent identity

A distinct, verifiable identity for an AI agent, separate from the human who uses it and from the application that hosts it. A strong agent identity lets a protected service know which agent is asking and on whose behalf.

Klyn perspective: An agent that borrows a person's credentials cannot be governed separately from that person. Klyn works alongside existing identity providers, including Okta, Microsoft Entra and Ping Identity, through supported integrations, and adds the agent-specific context those systems were not built to hold.

Agent owner

The named person or team accountable for an agent's purpose and access, and for responding when it misbehaves. Ownership is distinct from authorship, since the person who built an agent may not be the right person to answer for it.

Klyn perspective: Klyn works out who built each discovered agent and who should answer for it. High-risk actions can then pause for that named owner instead of a generic queue.

Attribution

Establishing who actually performed an action: the person, the agent acting for them, or both. Conventional logs often record only the account, which hides whether a human or an agent acted and what the agent was trying to do.

Klyn perspective: For investigation, Klyn attributes each action to the person who asked and the agent that carried it out, and sends the events to your SIEM and ticketing tools with the context an analyst needs.

Identity graph

A connected model of how agents relate to their owners, tools, data and the other agents they hand work to.

Klyn perspective: Klyn's Identity Graph relates agents, owners, tools and workflow handoffs where the source evidence supports those links. A relationship drawn in a graph is useful context; the protected service still needs trustworthy proof of who is asking.

Non-human identity (NHI)

Any identity used by software rather than a person, including service accounts, API keys, OAuth clients, certificates and workload identities. AI agents are a fast-growing category of non-human identity with one important difference: they decide what to do next.

Klyn perspective: Most NHI programs focus on secrets and credential hygiene. Agents also need controls on intent and action, because a well-managed credential can still be used for the wrong purpose.

Workload identity

A cryptographic identity issued to a running piece of software, such as a container or serverless function, based on where and how it runs instead of a stored secret. Workload identity reduces reliance on long-lived keys.

Klyn perspective: Workload identity tells a service which process is calling. For agents, the policy decision also needs the delegating person, the assignment and the specific action requested.

Authority and delegation

Authorization

The decision about whether an authenticated identity may perform a specific action on a specific resource. Authentication establishes who is asking; authorization decides what they are allowed to do.

Klyn perspective: For agents, authorization has to happen per action. Klyn checks protected operations against policy at the moment of action and records the decision, so permission to read a record never quietly becomes permission to change it.

Cascading revocation

Withdrawing authority from an agent or delegation so that the withdrawal also applies to every agent and grant that depends on it downstream. Without cascading revocation, a sub-agent can keep acting after its parent's access has ended.

Klyn perspective: Revocation is only real if the next protected operation fails. Teams using Klyn should test withdrawal on connected routes, including cached decisions, tokens and downstream sessions, and confirm the operation did not happen at the destination system.

Delegated authority

Permission that one party grants to another to act on its behalf within defined limits. In agentic systems, a person delegates to an agent, and that agent may delegate part of the task to another agent.

Klyn perspective: Delegated authority should get smaller as work moves downstream. Klyn's delegation controls give teams a place to define the limits of an assignment, such as resources, allowed operations, expiry and further delegation, and to trace each grant.

Delegation chain

The sequence of handoffs from the original human principal through one or more agents to the agent that finally performs an action. Each link should carry forward who authorized the work and the boundaries of that authority.

Klyn perspective: A helper agent investigating a bug has no reason to inherit deployment rights. Klyn helps teams follow the observed workflow through connected agents and tools, so authority can be checked at each protected step of the chain.

Just-in-time access

Granting access only when it is needed, for a specific task and a limited time, instead of keeping standing privileges. Just-in-time access shrinks the window in which a compromised or misdirected agent can cause harm.

Klyn perspective: Agents are well suited to just-in-time access because their tasks are discrete. Klyn favors authorization that is bound to a single request and expires quickly over broad, long-lived grants.

Least privilege

The principle that an identity should hold only the minimum access required for its task. For agents, least privilege applies to tools, data, operations and the right to delegate further.

Klyn perspective: Most agents run with broader access than their tasks require. Klyn's posture scoring ranks over-privileged agents by the damage they could do, so teams know which permissions to reduce first.

Mandate

A signed, verifiable statement of what a person authorized an agent to do, including constraints such as amount, merchant, scope or expiry. The Agent Payments Protocol (AP2) uses mandates as tamper-evident proof of user intent in agent-led payments.

Klyn perspective: A mandate helps only if it is verified. Teams adopting mandates should be clear about who signs and verifies them, and how each requested action is checked against the mandate before it executes.

Scoped, short-lived credentials

Credentials limited to specific resources and operations that expire after a short period. They reduce the value of a stolen token and limit what an agent can do if its instructions are manipulated.

Klyn perspective: In Klyn's execution integration, an allowed request receives a short-lived authorization bound to that request, which must be redeemed before the action runs. Reusing it, or changing the request, cannot turn one permission into another.

Token exchange

An OAuth pattern, defined in RFC 8693, for trading one security token for another with a different audience or scope. It is commonly used for on-behalf-of flows, where a service acts for a user with narrower rights than the original token carried.

Klyn perspective: Token exchange is a useful building block for agent delegation, provided each exchange narrows the scope. The receiving service still has to decide whether the specific action is allowed.

Policy and runtime control

Human-in-the-loop approval

A control that pauses a sensitive agent action until a designated person approves it. Effective approvals show the reviewer the exact action and its target, and the action waits for the decision.

Klyn perspective: High-risk actions pause for a named owner. In Klyn, an approval works once, for that exact action, and cannot be reused for a changed amount, recipient or target. Routine activity within scope proceeds under policy, which keeps approval fatigue down.

Kill switch

The ability to stop an agent or a provider immediately across the systems it can reach. A kill switch is only effective if it reaches the enforcement points the agent actually uses.

Klyn perspective: Suspension should be tested like any other control. Teams should confirm how quickly subsequent requests stop succeeding and whether any route around the enforcement point remains.

Point of action

The moment and place where an agent's request turns into an effect, such as a tool call or a transaction. Controls at the point of action can stop harm; controls that only alert afterwards cannot.

Klyn perspective: Alerts arrive after the damage. Klyn evaluates protected actions inline, before the tool call runs, and the decision rests on verified identity and delegated authority.

Policy decision point (PDP) and policy enforcement point (PEP)

Two roles from access control and zero trust architecture. The policy decision point evaluates a request against policy and returns a decision. The policy enforcement point sits in the path of the request and allows or blocks it based on that decision.

Klyn perspective: Klyn acts as a decision point for agent actions and integrates with enforcement points on the execution path. Coverage grows as more tools are integrated, and calls that bypass the integration need their own controls.

Policy simulation

Running a new or changed policy against real activity without enforcing it, to see what it would allow or block. Simulation lets teams tune controls before they interrupt legitimate work.

Klyn perspective: Klyn lets teams run a guardrail in simulation against real activity and see exactly what it would stop, then switch it to enforcement with confidence.

Runtime guardrails

Controls that evaluate an agent's actions while it runs and allow or stop them. Guardrails can apply to tool calls, data access, outputs and handoffs to other agents.

Klyn perspective: Klyn's runtime guardrails sit in the path of each protected action and decide to allow, ask, steer or block, weighing the whole session before the tool call runs.

Threats and risks

Agent goal hijack

An attack in which malicious content alters an agent's objectives or decision path, causing it to pursue an attacker's goal through otherwise legitimate actions. It is the first entry, ASI01, in the OWASP Top 10 for Agentic Applications.

Klyn perspective: A hijacked agent still has to act through tools. Policy that checks each protected action against the agent's assignment limits what a redirected goal can achieve.

Confused deputy

A classic security flaw in which a program with legitimate privileges is tricked into using them on behalf of a party that should not have that access. Agents are natural confused deputies because they act on instructions found in content.

Klyn perspective: The defense is to decide permission from the authenticated requester and the delegated authority, never from what the content says. A sentence inside a document or repository cannot grant permission.

Data exfiltration

The unauthorized transfer of data out of an organization. With agents, exfiltration can happen through ordinary-looking tool calls, such as sending an email or posting to a webhook.

Klyn perspective: Klyn's threat detection looks for exfiltration patterns in agent sessions and can block the action at an integrated enforcement point. Narrow assignments reduce how much data an agent can reach in the first place.

Excessive agency

Giving an agent more permissions or autonomy than its task requires, so that a mistake or manipulation leads to outsized harm. It appears in the OWASP Top 10 for LLM Applications.

Klyn perspective: Excessive agency is visible before anything goes wrong. Klyn's posture scoring highlights agents whose tools and permissions exceed their purpose.

Memory and context poisoning

Corrupting the information an agent relies on, such as stored memory or retrieved documents, so that it makes harmful decisions later. The poisoned content may be planted long before it is used.

Klyn perspective: Poisoned context changes what an agent wants to do. Enforcement outside the model's reasoning keeps those wants inside the boundaries of the assignment.

Privilege escalation

Gaining access beyond what was granted, for example by borrowing another agent's credentials or exploiting an over-scoped token. In multi-agent workflows, escalation often happens at handoffs.

Klyn perspective: Klyn's threat detection watches for privilege escalation during agent sessions. Delegation that narrows at each handoff makes escalation harder to begin with.

Prompt injection

Instructions hidden in content that an agent processes, such as a web page, email, document or tool description, which manipulate the model into ignoring its original instructions. Indirect prompt injection arrives through data the agent retrieves, so the user never sees it.

Klyn perspective: Detection helps, and Klyn looks for prompt injection in agent sessions. The stronger control is structural: the decision to allow an action rests on verified identity and delegated authority, so injected text cannot grant itself permission.

Rogue agent

An agent operating outside its intended purpose or authority, whether through compromise or gradual drift in its behavior. A rogue agent may look legitimate because it uses valid credentials.

Klyn perspective: Valid credentials are not evidence of valid intent. Klyn ties each action to its owner and to the policy decision that allowed it, so out-of-scope behavior stands out and can be stopped.

Tool poisoning

Planting malicious instructions or behavior in a tool that agents use, for example in an MCP server's tool descriptions or in a compromised dependency. Because agents read tool descriptions as guidance, a poisoned tool can steer them.

Klyn perspective: Klyn discovers MCP servers alongside agents and flags risky ones in its posture view. Treat tools as part of the agent's supply chain and review changes before agents rely on them.

Posture, testing and evidence

Agent observability

Visibility into what agents exist and what they do: their prompts, reasoning steps, tool calls, data access and outcomes. Observability is the foundation for investigation and audit.

Klyn perspective: Klyn's observability keeps a live inventory of agents with their owners and reach, and puts prompts and tool calls on one timeline, attributed to the person or the agent.

Agent red teaming

Adversarial testing of agents to find ways they can be manipulated into unsafe actions, such as leaking data or following injected instructions. A configuration review cannot show how an agent behaves under attack.

Klyn perspective: Klyn runs adaptive red-team scenarios against agents using synthetic data, so nothing real is put at risk. A confirmed finding becomes a guardrail, and Klyn replays the attack to show it now fails.

Agentic security posture management (Agentic SPM)

Continuous assessment of the security configuration and exposure of AI agents, including permissions, secrets, tools and connections, ranked by risk. It applies the posture management model used for cloud and SaaS to agents.

Klyn perspective: Klyn's AI security posture scoring covers over-privileged agents, exposed secrets and risky MCP servers, ranked by what to fix first.

Decision evidence

A reviewable record that links an agent action to the identity that requested it, the authority it relied on, the policy decision, any approval and the observed outcome. Decision evidence answers why something was allowed, in addition to what happened.

Klyn perspective: Klyn records each allow, ask and block with the agent, the owner and the policy that decided it. An allowed request is not the same as a completed change, so Klyn keeps authorization and execution outcomes distinct.

Protocols and frameworks

Agent2Agent (A2A) protocol

An open protocol that lets agents built on different platforms discover each other and exchange tasks. Agents can publish an Agent Card describing their identity and capabilities.

Klyn perspective: Interoperability makes delegation between agents routine, which raises the stakes on the delegation chain. Each handoff still needs bounded authority and a record of who authorized the work.

AI governance frameworks

Regulations and standards that set expectations for how organizations manage AI risk, including the EU AI Act, the NIST AI Risk Management Framework and ISO/IEC 42001. Auditors increasingly ask how they apply to agents.

Klyn perspective: Klyn's evidence and compliance capability maps recorded controls and evidence to frameworks including the EU AI Act, NIST AI RMF, ISO 42001, SOC 2 and GDPR.

Model Context Protocol (MCP)

An open protocol for connecting AI applications to external tools and data through MCP servers. For HTTP-based transports, the MCP authorization specification builds on OAuth 2.1, with a protected MCP server acting as an OAuth resource server. Authorization is optional in the protocol.

Klyn perspective: MCP standardizes how agents reach tools, which makes the tool call the natural place for policy. Klyn supports authorization for integrated MCP tool calls, alongside the native controls of the systems behind them.

MCP server

A service that exposes tools, resources or prompts to AI applications through the Model Context Protocol. MCP servers often hold credentials to business systems, which makes them a high-value target.

Klyn perspective: Klyn discovers MCP servers along with the agents that use them and ranks risky servers in its posture view. Every MCP server in use should have an owner and a reviewed set of tools.

NIST agent identity design tenets

Ten proposed principles from NIST for agent identity, authentication and authorization systems, published in October 2026 as a design direction, without the status of a formal standard. They cover distinct agent identities, bounded and verifiable delegation, independent decisions and trustworthy evidence.

Klyn perspective: We set out where Klyn's capabilities contribute to each tenet, and what an implementation still needs to demonstrate, in our analysis of NIST's design tenets.

OWASP Top 10 for Agentic Applications

A peer-reviewed list of the most critical security risks for autonomous AI agents, published by the OWASP GenAI Security Project in December 2025. It covers risks such as agent goal hijack, tool misuse, identity and privilege abuse, and rogue agents.

Klyn perspective: Many of the listed risks come down to agents acting with more authority than their task needs. The list is a practical starting point for threat modeling before an agent is deployed.

START WITH ONE AGENT

Put the vocabulary to work

Pick one agent your team already relies on and one action you need to control. Book a Klyn demo to see how that agent is discovered and which of its actions Klyn would check.

Book a Demo ↗

SECURE THE NEXT WORKFLOW

Ready to deploy agents with confidence?

Discover every agentic workflow, set clear authority, and retain decision evidence with Klyn.

Book a Demo