Box introduces Box Architect, a tool that designs, tests, and manages agentic workflows based on content already stored in Box. The company is also working on threat protection against agentic attacks. New security controls for agents are available immediately.
With Box Architect, teams can deploy AI agents across entire business processes. Users describe a process or goal in plain language. Architect then automatically determines the phases, decision points, and handoffs. In doing so, the tool draws on content and procedures already stored in Box. Existing permissions, classifications, and retention policies continue to apply at every step.
From design to production
Before a workflow goes live, teams can test it with real content and edge cases. They preview the agents’ output and routing and build in human approval at critical moments. After deployment, Architect monitors results and suggests improvements. Administrators can then standardize workflows across the entire organization.
Notably, users can choose a model for each step. They can select from models provided by Anthropic, OpenAI, Google, Nvidia, IBM, and others: more affordable models for simple tasks, and more powerful ones for complex reasoning. Box cites examples such as contract review at law firms, customer onboarding at banks, and claims processing at insurance companies.
According to CEO Aaron Levie, deploying AI across entire processes has often required specialized knowledge, but breakthroughs like OpenAI’s Astra and Anthropic’s Fable are changing that. “That’s why we’re building Box Architect – to help teams turn these advances in frontier AI into workflows that bring together agents, people, and content,” said Levie.
Agents as an attack vector
But more agents also mean a larger attack surface. Box is therefore working on threat protection against bulk exfiltration, session hijacking, account takeover, and mass file deletion. Unusual downloads must be detected more quickly and automatically restricted. Session binding and device-bound session credentials are designed to prevent stolen tokens from being reused on another device. Depending on severity, an automatic response follows, ranging from download restrictions to step-up authentication or session termination.
The agent security controls are already available. Classification-based policies determine which content agents can access. With MCP Guardrails, admins can restrict which tools agents are allowed to use, for example, for creating, moving, or sharing content externally. Threshold values alert teams as soon as an agent’s activity exceeds preset limits.
Box Architect will be rolled out as a beta to Enterprise Advanced customers later this quarter. The threat protection features will follow early next year for Enterprise Advanced and Shield customers. The agent security controls are generally available immediately for those same groups.
Tip: Box Agent now generally available as an AI engine for businesses