AI agent security glossary
18 terms a security or governance team meets when it starts on AI agents. Each has a short definition, the incident or standard behind it, and its sources.
Last reviewed . Guard0 Research.
Foundations
- Agentic AISoftware in which a language model chooses the steps toward a goal and carries them out with tools.
- AI agent securityProtecting AI agents, and the systems they can reach, from misuse and manipulation.
- Model Context Protocol (MCP)The open protocol that connects AI applications to external tools and data.
Attacks and failures
- AI red teamingStructured adversarial testing that finds how an AI system fails before an attacker does.
- Confused deputyA program with real authority that is tricked into using it for someone who has less.
- Excessive agencyAn agent that can do more than its task requires: too many tools, permissions or autonomy.
- MCP securityControlling the risks that MCP servers and clients add: open servers, held credentials, poisoned tools.
- Prompt injectionText that changes what a language model does, typed by a user or planted in content the model reads.
- Tool poisoningInstructions for the model hidden in the description, schema or output of a tool an agent can call.
Inventory and identity
- Agent identityThe credentials and attributes that let a system recognize an AI agent and decide what it may do.
- AI agent inventoryThe maintained list of every AI agent, model and AI tool: where it runs, what it reaches, who owns it.
- AI bill of materials (AIBOM)A machine-readable list of what an AI system is made of: models, datasets, tools and dependencies.
- Shadow AIAI tools, models and agents in use without the knowledge or approval of security and governance.
Governance and evidence
- AI agent observabilitySeeing what an agent did while it ran, and comparing that with what it normally does.
- AI governanceThe policies, roles and controls that decide which AI runs, who answers for it and how that is evidenced.
- AI security posture management (AI-SPM)Tooling that discovers AI systems, checks their configuration and access against policy, and tracks the risks.
- Decision recordThe kept trail of an agent run and the decisions about it: data seen, tools called, policy, approver.
- Human of RecordThe named person who answers for an AI agent: what it can reach, what it does, what was decided about it.
Missing a term, or see something out of date? Write to hello@guard0.ai.
Find the agents in your own company.
Start free: discover the AI on three people's machines, register one AI system and connect one repository or AWS account.