Brex developed a novel approach to AI agent security
Learning security policies at the network layer enhances AI safety
1Limitations of traditional security measures
Brex realized AI agents couldn't be effectively managed with traditional SDK permissions and model constraints. Agents requiring real credentials like API keys and service accounts could bypass existing security frameworks, necessitating a new approach.
2Revolutionary solution at network layer
Brex's CrabTrap platform is an open-source proxy intercepting all network traffic. It enforces policy rules using an LLM-based 'judge' for HTTP/HTTPS requests. This language- and framework-agnostic approach provides exceptional flexibility.
3Dynamic policy creation
The system observes agents' actual behavior to create policies beyond static rules. By analyzing historical network traffic, it generates natural language policies. This method proves significantly more effective than starting from scratch.
4LLM judge mechanism
CrabTrap employs an LLM-based decision mechanism for unfamiliar request patterns. This judge activates for only about 3% of requests in mature agents, minimizing performance impact while maintaining security.
5Policy evaluation system
Brex built an evaluation system testing policy changes before deployment. Comparing against historical audit logs, it measures impact within minutes. Results can be filtered by method, URL, and decision status for precise analysis.
6Real-time feedback loop
The system operates on complete audit trails stored in PostgreSQL. For consistently denied resources, it can trigger human intervention or suggest policy updates, creating a closed-loop improvement process.
7Performance optimizations
Despite concerns about LLM judge latency, the system outperformed expectations. The judge's infrequent activation and agents' rapid learning to avoid repeat requests contributed to efficient operation.
8Multi-layered security philosophy
Brex views the transport layer solution as complementary to existing measures. Their 'security by layers' approach positions this innovation as filling an underinvested gap in comprehensive protection.
9Industry implications
Brex's innovation may set new standards in AI agent security. Shifting from traditional permissions to a learning, adaptive network control plane represents a paradigm shift for IT leaders managing AI deployments.