Brex is overhauling how it polices its AI systems, shifting from static code constraints to live network traffic monitoring. The company is using an open-source LLM proxy to watch AI interactions in real time, a move that could reshape enterprise security practices.
A Shift From Code to Live Oversight
Traditional AI governance often means embedding rules directly into the model's code or relying on post-hoc audits. Brex is taking a different route. By monitoring network traffic, the company can see every request and response as it happens, catching policy violations or anomalies the moment they occur. That's a fundamental change from the old model, where problems might only surface after the fact.
The shift is notable because it treats AI governance as a runtime concern, not a design-time one. Instead of hoping the model behaves, Brex is watching what it actually does.
Why Network-Level Monitoring
Network-level monitoring gives a complete picture of AI activity. It doesn't matter whether the model is internal or external, cloud-hosted or on-premises. If traffic flows through the network, it can be inspected. That means Brex can detect data exfiltration attempts, unauthorized access, or even subtle prompt injection attacks that might slip past code-level checks.
Real-time monitoring also allows for immediate response. If something looks wrong, the system can block the request or alert the security team before any damage is done. That's a level of agility that static rules simply can't match.
The Open-Source Proxy Approach
Brex is relying on an open-source LLM proxy to handle this traffic inspection. The proxy sits between users and the AI models, logging every interaction and applying policy filters. Because it's open source, the code is transparent and can be audited by anyone. That's a deliberate choice—it means the governance layer itself is subject to scrutiny, not just the AI models it oversees.
Open-source also brings flexibility. Brex can modify the proxy to fit its specific needs, and the broader community can contribute improvements. It's a collaborative approach to security, one that avoids the black-box problem of proprietary monitoring tools.
Redefining Enterprise Security
This move could have ripple effects across the industry. If Brex's approach proves effective, other companies may follow suit, adopting network-level monitoring as a standard part of their AI governance stack. The idea is that security shouldn't be bolted on after deployment—it should be woven into the infrastructure itself.
That's a bold claim, but the logic is sound. AI systems are becoming more complex, and the attack surface is growing. Monitoring at the network layer catches things that code review never will. It's a proactive stance, not a reactive one.
Brex hasn't said when the system will be fully deployed across all its operations. The bigger question is whether this approach scales beyond a single company—and whether regulators will start expecting it as a baseline. For now, Brex is betting that watching the traffic is the best way to stay ahead.
