CrabTrap: Brex's AI Agent Network Guardian

Brex has engineered an internal platform named CrabTrap to tackle the security vulnerabilities associated with widespread AI agent adoption in enterprise environments. Recognizing that traditional guardrails were insufficient for agents wielding real credentials, Brex created CrabTrap to act as an HTTP/HTTPS proxy. This system intercepts all network traffic, scrutinizes requests against established policy rules, and utilizes a novel LLM-as-a-judge to make approval or denial decisions. The core strategy shifts agent governance from SDK-level permissions to a centralized network control plane that learns from real-world agent behavior.

The CrabTrap platform operates at the transport layer, making it agnostic to specific agent frameworks, languages, or APIs. By setting environment variables, all outbound agent requests are routed through CrabTrap for inspection. This approach is part of Brex's layered security philosophy, focusing on an underinvested enforcement point. The LLM judge is judiciously applied, typically engaging for fewer than 3% of requests that deviate from established patterns, thus minimizing latency. Brex has addressed potential prompt injection risks by structuring all user-controlled content within requests as escaped JSON objects.

Brex has released CrabTrap as open-source, inviting community contributions to enhance its capabilities. Future developments may include deeper authentication features like SSO and RBAC, escalation workflows for agents requiring extra permissions, and policy recommendations derived from denial patterns. The platform's success has bolstered organizational confidence in deploying AI agents broadly, transforming CrabTrap from a mere enforcement tool into a valuable discovery mechanism for optimizing agent performance and tightening overall security.