The attack surface moved inside the agent. So did Arcjet.

Arcjet launched Guards, a runtime security feature built into its SDK for AI agent workflows.

It inspects untrusted input inside tool handlers, queue consumers, and workflow steps, where perimeter tools like WAFs cannot see.

The feature is designed to stop prompt injection, malicious tool output, hidden instructions in fetched content, PII leakage, and runaway token costs.

Arcjet says moving enforcement into the code path gives developers better control as more application logic shifts to agents and background jobs.