AI agent security glossary
Twenty-eight terms define the vocabulary of AI agent security, from agent runtime, agent registry and non-human identity to kill switch, blast radius, memory poisoning and goal hijacking. Each entry gives a short definition and one sourced example read on 16 Nov 2026.
Facts read on .
28 terms
- A2A protocol: The A2A protocol, short for Agent2Agent, is an open specification that lets agents built on different frameworks discover each other's capabilities and exchange tasks and messages.
- Access request workflow: An access request workflow is the path by which a user or agent asks for something policy blocks, and an administrator approves or denies it.
- Agent credential vault: An agent credential vault is a managed store for the keys, tokens and secrets an agent needs, kept outside the agent so it never holds them in plain view.
- Agent framework: An agent framework is a code library that supplies the building blocks for an agent: model calls, tool definitions, memory and control flow.
- Agent lifecycle: The agent lifecycle is the sequence from creating an agent through approval, change, monitoring and retirement.
- Agent observability: Agent observability is the collection of prompts, responses, tool calls and outcomes that show what an agent did and why.
- Agent orchestration: Agent orchestration is the logic that decides which agent or tool acts next, passes state between steps and handles retries.
- Agent registry: An agent registry is a central list of the agents, tools and skills in use, with an owner and an approval status for each.
- Agent runtime: An agent runtime is the software environment in which an AI agent executes its loop of model calls, tool calls and state handling.
- Agent sprawl: Agent sprawl is the spread of AI agents across platforms, SDKs and employee machines without a central record of who owns them or what they can reach.
- Agent trace: An agent trace is the ordered record of one agent session: the prompts, the tool calls, the results and the downstream actions they caused.
- AI browser: An AI browser is a web browser with a built-in AI assistant that can read pages, fill forms and take actions for the user.
- Autonomy level: An autonomy level states how much an agent may do on its own, from suggesting actions to acting without any approval.
- Blast radius: Blast radius is the amount of data, systems and other agents that one compromised or misbehaving agent can reach.
- Goal hijacking: Goal hijacking is an attack in which hostile content changes what an agent is trying to do, so it pursues the attacker's aim instead of the assigned task.
- Guardrail policy: A guardrail policy is a written rule that maps a detected condition to an action, such as alert, block, redact or log.
- Just-in-time access: Just-in-time access grants an identity a permission only when a task needs it and removes it when the task ends.
- Kill switch: A kill switch is a control that stops an agent, cuts its access or isolates it when risk is high.
- LLM gateway: An LLM gateway is a proxy that sits between applications or agents and the AI models they call.
- Memory poisoning: Memory poisoning is the planting of false or hostile instructions in the long-term memory an agent keeps between sessions.
- Model routing: Model routing is the practice of sending each request to a model chosen by cost, task or policy.
- Multi-agent system: A multi-agent system is a set of AI agents that divide a task and pass work, context and results to each other.
- Non-human identity: A non-human identity is an account, key, token or certificate used by software or a machine instead of a person.
- OAuth scope: An OAuth scope is a label in an OAuth request that names the kind of access a token grants, such as read-only access to one resource.
- Permission boundary: A permission boundary is a ceiling on the permissions an identity can receive, set apart from the policies that grant them.
- SaaS agent: A SaaS agent is an AI agent that runs inside a software-as-a-service platform, using that platform's data, tools and identity model.
- Unbounded consumption: Unbounded consumption is runaway use of model calls, tokens or compute, either by accident or on purpose.
- Workspace agent: A workspace agent is an AI assistant or agent embedded in productivity and collaboration software, with delegated access to a user's documents, mail and chats.