Antoine Brossault
“My AI agents are the new target for attackers”
Antoine Brossault · Loire Room · 15:20 · 40 min
“My AI agents are the new target for attackers”
Security, sandboxing and least privilege at company scale
By 2026, everyone has an agent. Almost nobody has governance.
At Cloudflare, the whole company was equipped in under a year, and merge velocity doubled. What made the difference was not the model we picked. It was the architecture around it.
This talk explains why we built our own AI stack, and the decisions that followed: centralizing every model call behind a single gateway, moving security to the network layer instead of into each tool, governing access to internal data by identity, and running agent-generated code in an isolated environment.
We’ll talk architecture, results and terminal, and we’ll stand behind one opinion: the harness and the model are two separate things.
We’ll cover:
We’ll illustrate this with Cloudflare OS, our open source harness, and in particular its Gatekeepers mechanism. They provide stronger security than a classic MCP integration: each agent or application gets only the capabilities it strictly needs, actions are logged, and sensitive operations can require human approval.
The project and the details of this architecture: