Arcjet Introduces Guards for Security Inside AI Agent Workflows

Arcjet announced in a press release a new capability called Guards, designed to enforce security directly inside AI agent workflows, queue consumers, and other application code that does not run behind an HTTP request. The feature extends Arcjet's runtime security model into long-running systems where traditional perimeter tools cannot inspect untrusted input.

Guards allow developers to define and enforce policies at the points where untrusted input enters an application. This includes tool calls, queue messages, and workflow steps. The enforcement happens within the application's context, such as identity, session state, or business logic, rather than at a network boundary.

The new capability integrates with Arcjet's existing application-layer security model. Developers can define rules within the same codebase as the application, ensuring that protection is reviewed and deployed alongside the feature itself. Guards can detect prompt injection, block sensitive data before it reaches external models, enforce user token budgets, and validate input in background jobs.

Guards work alongside Arcjet Shield and other existing protections, covering both public endpoints and internal code execution paths. The feature is available now through Arcjet's JavaScript and Python SDKs for both existing and new users.