--- title: "Securing AI Agents: Why Brex Watches the Network" description: "Discover how Brex secures autonomous AI agents at the network layer instead of relying on code. Learn to build impenetrable guardrails for your brand." canonical: https://epinium.com/en/blog/brex-ai-agents-network-guardrails/ lang: en date: 2026-08-12T05:09:00 --- **Executive summary** - **The paradigm shift:** Brex CEO Pedro Franceschi revealed at VB Transform 2026 that securing AI agents requires watching the network, not the code. - **The real impact:** Through a radical "immune system" approach, an independent LLM now governs which network requests Brex's internal agents can make, intervening in just 2% of cases. - **The takeaway for brands:** To scale autonomous AI without risking data leaks, CTOs and COOs must stop trying to build perfect prompts and instead build impenetrable network firewalls. You are standing at the edge of a cliff. Your competitors are deploying AI agents that analyze market trends, adjust ad campaigns, and negotiate with suppliers. Your team, meanwhile, is still drowning in manual spreadsheets. You want to unleash autonomous AI. But your CTO is terrified. What if the agent hallucinates and leaks proprietary pricing data to a public server? What if it deletes a core database? Here is where the majority of tech leaders get it completely wrong. They try to control the AI by writing longer prompts or adding more code constraints. They treat the agent like a rigid software script. This is a massive mistake. At VB Transform 2026, Brex CEO Pedro Franceschi shattered this exact myth. His [blueprint for enterprise AI deployment](https://venturebeat.com/orchestration/brex-assumes-its-ai-agents-could-do-anything-so-it-watches-the-network-not-the-code) does not rely on micromanaging the model. It relies on assuming the agent will eventually go rogue, and building the guardrails at the network layer instead. ## The "immune system" mindset and why perfect code is a trap If you manage a consumer goods portfolio, you know the daily chaos of [What Does A Brand Manager Do](/en/blog/what-does-a-brand-manager-do/) — juggling inventory, ad spend, and competitor analysis across multiple channels. AI agents are the obvious solution to this operational nightmare. But deploying them securely remains the ultimate bottleneck. Franceschi argues that trying to constrain a Large Language Model with code is a losing battle. The nature of these models is probabilistic. They will find unexpected, sometimes dangerous paths. Brex's playbook is radical but incredibly pragmatic. Let the agents operate freely inside a sandbox, but tightly control their access to the outside world. > **33%** — The proportion of companies that have successfully scaled AI across their organization, highlighting the massive execution gap in enterprise deployments. [Source: McKinsey 2025](https://www.mckinsey.com/capabilities/quantumblack/our-insights/the-state-of-ai) Instead of relying on static policy PDFs or endless prompt engineering, Brex enforces its security policy at the network perimeter. An independent, specialized LLM acts as the gatekeeper. This overseer decides exactly which network requests the working agents are allowed to execute. Surprisingly, this gatekeeper only has to step in about 2% of the time. Operations stay fast. Zero unauthorized data exfiltration occurs. FREE SESSION **Stop guessing how to deploy AI securely.** Take control of your infrastructure today. [Discover Transform →](/en/transform/) free 30-min diagnostic ## Legacy constraints vs. The network immune system To understand why this matters for your brand, look at the fundamental differences between how companies used to deploy AI and how pioneers are doing it today. | Feature | Legacy Prompt-based Security | Network-level Guardrails (Brex approach) | | --- | --- | --- | | **Core philosophy** | Constrain the AI's thoughts | Constrain the AI's actions | | **Vulnerability** | High risk of prompt injection | Extremely resilient to injection | | **Deployment speed** | Slow (endless edge-case testing) | Fast (safe execution sandbox) | | **Governance layer** | Inside the model code | Independent gatekeeper LLM | | **Enterprise fit** | Fails at scale | Designed for autonomous workflows | ## A reality check for COOs and Marketing Directors Many brands abandon their automation projects because they fail at the context and security layers. You can read more about this exact failure point in our breakdown of [Why Enterprise Ai Agents Fail Agentic Context Layer](/en/blog/why-enterprise-ai-agents-fail-agentic-context-layer/). They give an open-source agent like OpenClaw access to their Amazon Seller Central without a proper safety net, panic at the first hallucination, and pull the plug. If you are a COO or a marketing director, the Brex approach should be your new operational blueprint. You do not need a perfectly behaved model. You need a secure execution environment. The companies winning this race treat their AI agents like brilliant but reckless interns. You don't trust them with the master keys to the corporate bank account. You give them a sandbox, a clear task, and a strict supervisor. > **Epinium data:** Brands implementing network-level guardrails deploy AI agents 3x faster than those relying solely on prompt-based constraints. ## Databricks and the "Innocent Until Combined" philosophy Brex is not alone in this architectural shift. Enterprise data giants like Databricks are pushing in the exact same direction. They are shipping contextual policies that block "the lethal trifecta": sensitive internal data combined with untrusted external content and outbound network calls. This is the exact philosophy manufacturers must adopt to protect their supply chain data while still moving at the speed of modern commerce. [Gartner reports](https://www.gartner.com/en/newsroom) that 40% of enterprise apps will feature AI agents by 2026. You cannot afford to sit on the sidelines just because you are afraid of the technology. It is time to shift your focus. Stop asking if the agent can do the task. Start asking if the agent is operating in an environment where its inevitable mistakes cannot hurt you. ## Frequently Asked Questions ### What is the main challenge of deploying AI agents in the enterprise? The biggest hurdle is security and governance. Traditional software relies on predictable code, but AI agents are probabilistic. They can hallucinate or execute unintended actions, making it dangerous to give them unrestricted access to sensitive brand data or external networks. ### How does Brex secure its internal AI agents? Instead of trying to write perfect code or prompts to constrain the AI, Brex assumes the agent might compromise itself. They enforce security at the network layer. A secondary LLM acts as a gatekeeper, monitoring and approving all network requests made by the primary agents. ### Why do prompt-based security measures fail? Prompt engineering is easily bypassed through direct or indirect prompt injections. Relying solely on prompts to keep an AI agent secure is like asking a thief to promise not to steal. The guardrails must exist outside the agent's control, preferably at the infrastructure and network level. ### What does this architectural shift mean for brand managers? It means you must change your deployment architecture. You can safely deploy autonomous agents to handle inventory, marketing, or competitor analysis as long as you sandbox their environment and monitor their network activity, rather than obsessing over the model's internal logic. ### How often does the network gatekeeper LLM need to intervene? According to Brex's presentation at VB Transform 2026, their gatekeeper LLM only needs to block or intervene in about 2% of the total network requests, ensuring that the agents operate quickly without compromising enterprise security. ## The new standard for AI orchestration The era of isolated AI chat interfaces is officially over. Agentic workflows are here, and they are hungry for data, APIs, and network access. Franceschi's revelation is a massive wake-up call. You cannot prompt your way to a secure enterprise. You must build it into the bare metal of your network. TRANSFORM BY EPINIUM **Scale your brand faster than the competition.** See how our experts can rewire your workflows. [Book free diagnostic →](/en/contact-transform/) free 30-min diagnostic