The AI agent security glossary

The terms that matter when you're shipping agents into production. No fluff, just definitions.

• Prompt injection

  Malicious instructions embedded in content an AI agent processes — a document, email, webpage, or API response. The agent reads it, treats the instruction as legitimate, and executes it using real credentials. No malware required. Just text.

• Memory poisoning

  Malicious content stored in an AI agent's long-term memory or RAG database. When the agent retrieves it in future sessions, the embedded instructions reactivate — causing malicious behavior long after the original attack. Current detection rates are below 20%.

• AI supply chain attack

  Compromising a third-party dependency an AI system relies on — a model, plugin, MCP server, or agent framework. Unlike traditional supply chain attacks, the payload can be the instructions the agent receives at runtime, not just the code it runs.

• Agent attack vectors

  The main surfaces through which autonomous agents are compromised: prompt injection, tool hijacking, memory poisoning, goal hijacking, multi-agent chain attacks, and supply chain compromise. Each targets a different layer of how agents perceive, reason, and act.

• Runtime policy enforcement

  Intercepting every tool call, API request, and agent action before it executes — and blocking anything not explicitly permitted by policy. The only control that operates after an agent has decided to act. Allowlist-only: everything not permitted is blocked by default.

• AI agent red teaming

  Adversarial testing of AI agents before production — simulating prompt injection, tool abuse, goal hijacking, jailbreaks, and multi-step attack chains. Targets the agent's reasoning and decision-making, not its network exposure. Distinct from traditional penetration testing.

• MCP security

  Securing agents that connect to external tools via the Model Context Protocol. MCP servers can inject hidden instructions, impersonate trusted tools, or redirect agent actions — all executing with the agent's full permissions. 492 MCP servers were found publicly exposed with no authentication in 2026.

• AI agent security

  Controlling, monitoring, and enforcing what autonomous AI agents are allowed to do at runtime. Covers access control, policy enforcement, red teaming, and audit logging across every layer of an agent's execution path. Distinct from AI safety, which focuses on model alignment.

• AI agent identity

  Treating AI agents as authenticated entities with explicitly scoped permissions — the same way you govern service accounts or human employees. Only 21.9% of organizations do this. The rest run agents with shared credentials and no agent-level access controls.