The AI security glossary.

Agent guardrails are the controls that keep an AI agent's behavior inside safe, intended boundaries, deciding what it may do, what it must ask about, and what it must never do. As agents gain autonomy and inherit human levels of trust, guardrails are what let organizations grant that autonomy without losing control.

Agentic AI security is the discipline of securing autonomous AI agents and the entire environment they touch: the agents themselves, the shadow AI nobody approved, the skills and MCP servers they pull in, the data they move, and the actions they take at runtime. It is broader than watching what an agent does. It is control over everything that can influence an agent and everything an agent can reach.

AI agent drift is when an agent's behavior moves away from what it was intended to do, either within a single task or gradually across many. It is one of the clearest early signals that an agent is being manipulated, malfunctioning, or operating outside its purpose.

AI agent observability is the ability to see what an AI agent is doing, why, and with what result, across its full execution. As agents take on multi-step, autonomous work, observability is what turns an opaque black box into something teams can debug, secure, govern, and account for. It also ties into the ability to see the entire agentic estate, including Shadow AI.

AI Detection and Response (AIDR) is the security discipline focused on detecting risky or malicious behavior by AI systems, especially autonomous agents, at runtime, and responding fast enough to prevent or contain the damage.

An LLM jailbreak is a technique for getting a large language model to ignore or work around the behavioral safeguards it was trained and configured to follow.

MCP security is the practice of governing and protecting the Model Context Protocol connections that let AI agents reach external tools, data, and systems, along with the skills and other components agents load alongside them. As MCP becomes the default way agents extend themselves, it has also become one of the fastest-growing attack surfaces in the enterprise.

MCP Tool Poisoning is a security threat in which malicious or manipulated instructions embedded within a Model Context Protocol (MCP) tool influence an AI agent’s behavior.

MCP is an open protocol that lets AI models connect to external tools and data sources through a standard interface — and, in doing so, creates a new supply chain to secure.

Prompt injection is an attack that manipulates a large language model by feeding it crafted input, causing it to ignore its intended instructions and follow the attacker's instructions instead. It is consistently ranked as the top security risk for LLM applications.

Shadow AI is the use of AI tools, models, and agents inside an organization without the knowledge or approval of IT and security. It is the AI equivalent of shadow IT, and because agents act rather than just answer, its risks run deeper.