HTTP Proxy Security
person concept tool org talk claim — click a node to jump to its page; hover an arrow for the relation
A network-layer security approach for AI agents in which all traffic crossing an agent's network boundary is routed through an HTTP proxy, enabling policy enforcement and monitoring without modifying the agent itself.
Core Concept and Motivation
Pedro Franceschi argues that controlling AI agent behavior at the application level is insufficient, and that "the only way to actually do something about it was to do something in the network layer." → This framing positions HTTP proxy security as a foundational architectural requirement rather than an optional add-on — the implication being that attempts to secure agents through prompt engineering, application-level filters, or other higher-level mechanisms alone will fail to provide adequate guarantees.
Implementation via Crab Trap
Crab Trap is a concrete tool that instantiates this approach by HTTP-proxying "the entire network boundary of an agent." → By intercepting all outbound network calls, the proxy can apply policy enforcement — including LLM-as-judge evaluation of traffic — uniformly across any agent regardless of its internal implementation.
Strategic Rationale for Enterprise AI
The broader claim advanced by Franceschi is that securing agents at the network layer is the correct architecture for enterprise AI deployments specifically because it decouples security from experimentation: organizations can adopt aggressive AI agent usage while maintaining guardrails, since the enforcement point sits outside the agent rather than within it. → This mirrors established enterprise network security practice (e.g., TLS inspection proxies, API gateways) applied to the AI agent context.
Points of Disagreement
No speakers in the current graph dispute this approach. The advocacy for HTTP proxy security comes exclusively from Franceschi and the architectural claims associated with Crab Trap.