Brex builds AI agent control layer at network level, not in rules

Brex, the fintech company, has built an open-source solution called CrabTrap to address a blind spot in how AI agents are secured at enterprise scale. The problem emerged as agentic frameworks like OpenClaw became widely adopted: while these frameworks enable agents to take actions autonomously, they lack proven mechanisms to govern agent behavior when agents have access to real credentials.

The core issue is that agents need genuine API keys, OAuth tokens, and service accounts to function effectively—they are not merely simulated tools but real gateways to live systems. Traditional guardrails, whether embedded at the SDK level or baked into model training, proved insufficient. Brex co-founder and CEO Pedro Franceschi explained to VentureBeat: "What we noticed was that the network layer was an untapped enforcement point. Every request an agent makes is an opportunity to intercept, reason about, and make a policy decision."

CrabTrap works by operating as an HTTP/HTTPS proxy that sits between agents and the systems they request. It intercepts every outbound request, examines the request against policy rules, and uses a LLM-as-a-judge—an AI system trained to evaluate compliance—to determine whether the request should be approved or denied. This approach centralizes control at the network boundary rather than trying to constrain the agent itself.

Franceschi's recommendation to IT leaders is that agent governance should shift away from SDK-level permissions and model guardrails toward centralized network control. By enforcing policy at the network layer, organizations can maintain audit trails and real-time oversight of agent activity, regardless of what code the agent runs or how it was trained. For enterprises deploying agents with real credentials, this architectural change may become standard practice.

The article identifies a structural gap in how AI agents are currently managed. OpenClaw and similar agentic frameworks have become widely adopted, but they operate without proven safeguards at enterprise scale. The challenge is that agents need genuine credentials—API keys, OAuth tokens, service accounts—to be useful, yet traditional guardrails (permissions embedded in the agent's code or baked into the model) cannot effectively limit what agents do once those credentials are in hand.

Brex's insight is architectural: the network layer becomes the enforcement point. Every request an agent makes traverses the network, creating an observation and control opportunity that is independent of the agent's code or training. By placing a proxy at this chokepoint and having it evaluate each request against policy rules using an LLM-as-a-judge, Brex sidesteps the limitations of SDK-level permissions and model-level guardrails. Franceschi's message to IT leaders is that this shift—from application-layer controls to network-layer controls—should become the standard practice for agent governance.

No comments yet. Be the first to share your thoughts!