Crab Trap
person concept tool org talk claim — click a node to jump to its page; hover an arrow for the relation
Crab Trap is an open-source agent security tool built and used by Brex in production, developed under the direction of Pedro Franceschi. It is designed to secure AI agents at the network boundary before their actions reach external systems.
Architecture and Approach
Crab Trap operates by proxying all network traffic from an agent: it "HTTP proxies the entire network boundary of an agent", intercepting outbound requests before they are fulfilled. This positions the tool as a runtime enforcement layer rather than a static policy definition.
LLM-as-Judge for Policy Enforcement
At the core of Crab Trap's enforcement mechanism is an LLM as a judge pattern. According to Franceschi, the system uses "an LLM as a judge and the LM determines is this request something that should be approved or not based on the policy for what that agent should be doing". This means agent behavior is validated dynamically against a stated policy rather than through hard-coded rules.
Provenance and Deployment
Franceschi describes Crab Trap as the production mechanism Brex actually relies on for agent security: "the way we use to secure agents at Brax in production". It was open-sourced approximately two months prior to the talk, making Brex's internal production approach publicly available to other teams building agentic systems.