Agentic infrastructure is a system that can interpret user input, reason over context, and take actions through connected tools or services. In search and assistant products, that can mean going beyond returning results to drafting messages, filling forms, changing settings, or triggering external workflows. The key shift is from information retrieval to delegated execution.
From a security perspective, this widens the attack surface. If the system acts on behalf of a user, attackers may try prompt injection, malicious web content, or deceptive instructions to influence its decisions. Defenders need tight authorization, least-privilege tool access, explicit confirmation for risky actions, strong provenance checks, and audit logs that show what was read, decided, and executed. Without those controls, a helpful agent can become an untrusted operator.