Box AI Agents — agentic threat model

Integrates with external top LLM providers via Box AI Studio. Risks include adversarial prompt injection bypassing document boundaries, and potential data leakage to third-party LLM APIs if data-sharing agreements or zero-data-retention policies are misconfigured.

Processes highly sensitive unstructured enterprise documents (HR, Legal, Finance) using permissions-aware search and metadata extraction. Risks include indirect prompt injection via malicious document uploads designed to exfiltrate data or poison the search index.

Orchestrates workflow automation and document approvals. Risks include insecure tool execution where an attacker manipulates the agent's decision-making flow to trigger unauthorized approvals or execute unintended workflow steps.

Not certain from the listing — Hosted within Box's enterprise cloud infrastructure. General risks include container breakout, API vulnerabilities in Box AI Studio, and insecure transit of document payloads to external LLM providers.

Not certain from the listing — Likely relies on Box's standard enterprise logging and audit trails. General risks include lack of specialized LLM alignment monitoring, drift in extraction accuracy, and silent failures in compliance checks.

Strong focus on permissions-aware access control, compliance, and governance. Risks include complex RBAC misconfigurations allowing horizontal privilege escalation across document repositories, or failures in mapping user identities to LLM context windows.

Allows creation of multiple tailored agents (HR, Finance, Legal) via Box AI Studio. Risks include cross-agent trust exploitation where a compromised HR agent accesses sensitive financial agent workflows or shares unauthorized context.