A panel of enterprise security leaders including CISOs and AI startup founders discussed how to govern AI agents without blocking developer productivity. Key themes included isolating agents in sandboxes, keeping humans in the loop for dependency selection, treating laptops as production environments, and securing MCP server usage. Docker's approach centers on disposable MicroVM-based sandboxes and an MCP Gateway that enforces authentication and logging for every tool call. Panelists recommended layered security, blessed image repositories, minimum release age policies for dependencies, and a training-wheels approach to building agent skills before handling sensitive data.
Table of contents
Isolate, control, observeNanoClaw—a personal AI agentLaptops as the new prodPortability—from laptop to cloudWhen agents build the supply chainKeeping humans in the loopA training-wheels approachLayer security to limit the blast radiusMCP—the new shadow IT?Making it safe and easy to run MCP serversA once-in-a-generation opportunity—and challenge295 Impressions