AI agent security glossary.

Software that uses an AI model to plan and take multi-step actions toward a goal, typically by calling tools, APIs, browsers, code interpreters, or other agents with some degree of autonomy.

The discipline of protecting systems from risks created by autonomous AI agents, covering identity, authorization, tool use, memory, runtime behavior, and interactions between multiple agents.

A synonym for AI agent security that emphasizes the agentic properties — planning, tool use, and autonomy — that distinguish agents from standalone language models.

A system in which several AI agents interact, either through central orchestration with fixed roles or through distributed coordination, to complete tasks no single agent performs alone.

A population of AI agents whose behavior is coordinated through distributed, local, or emergent interaction rather than a single controller, producing group behavior that differs from any individual agent.

Harmful activity performed by multiple autonomous agents whose combined behavior is more capable, adaptive, or difficult to detect than any individual action. Coordination may be explicit or emerge through shared resources.

Activity across a population of machines that begins to behave like a single coordinated system — for example parallel exploration followed by converging actions.

The spread of a capability, discovery, technique, or behavioral change from one agent to others across a machine population, established through time-ordered evidence.

Any shared resource — file, repository, cache, queue, memory store, or environmental side effect — that carries information from one agent to another, even without an explicit message.

Indirect coordination in which agents communicate by modifying a shared environment that other agents later observe and react to. The term originates in the study of social insects.

Multiple agents reading or writing the same resource. Co-access alone is common and benign; it becomes evidence of communication only when a distinctive change measurably influences later behavior.

An agent that connects otherwise separate subgroups of a machine population, often the path through which capabilities or information propagate between them.

An agent whose position and timing in a population graph suggest it may be directing or synchronizing others. It remains a candidate until supporting evidence establishes the role.

Coordinated attempts by multiple agents against identity, policy, or resource boundaries — for example many agents probing the same permission from different directions.

An AI agent operating outside its intended behavior, objectives, or authority, whether through compromise, goal hijacking, misalignment, or emergent drift. Listed as ASI10 in the OWASP Top 10 for Agentic Applications.

A fault or compromise in one agent that spreads through connected agents, tools, and trust chains, amplifying impact. Listed as ASI08 in the OWASP Top 10 for Agentic Applications.

Inwom's trust principle: agent transcripts are claims, trusted infrastructure activity is evidence, inferred relationships stay labelled as inferred, and AI reasoning never becomes authority.