Brex Develops Network-Layer Security for AI Agents

Fintech company Brex has created a new network-level security layer named CrabTrap to protect AI agents, such as OpenClaw, deployed in production environments. According to CEO Pedro Franceschi, traditional security models are insufficient for securing agents with code execution capabilities.

Franceschi presented the solution at VB Transform 2026, emphasizing that the enterprise deployment of AI agents necessitates a mindset shift. The goal is to create entities that can genuinely collaborate with human workers, akin to virtual employees.

Brex's security team initially rejected the idea of allowing agents like OpenClaw, which possess code execution abilities, access to internal systems. To address this, Brex developed CrabTrap, an HTTP proxy that monitors all outbound network traffic between the agent and the internet. It employs a large language model (LLM) to judge whether the traffic aligns with approved policies.

The system processes low-risk actions instantly via pre-approved rules. High-risk actions, such as sending emails, are flagged for LLM evaluation. Franceschi stated that only about 2% of requests experience LLM latency, with the LLM judging network traffic based on its inherent understanding.

Brex tested CrabTrap with a virtual recruiter named "Jim." When Jim attempted an action outside policy, CrabTrap initiated a human review via Slack. Franceschi acknowledged CrabTrap as a costly solution, but its development provided Brex with an advantage in AI adoption.