Europe 2026
What if the network was the sandbox? — Remy Guercio, Tailscale
Overview
This talk proposes a shift in how we conceptualize sandboxes for AI agents, suggesting the network itself can serve as the sandbox. Instead of relying on traditional methods like API keys or OAuth within a VM or container, the approach leverages network-level identity and permissions. This allows for granular control over agent access and behavior directly at the network layer, enabling more secure and observable AI agent deployments.
Who should watch
- AI Engineers
- Product Managers building AI-powered features
- Developers working with AI agents and sandboxing
- Security professionals concerned with AI agent access control
- Anyone interested in novel approaches to network security and AI infrastructure
Key takeaways
- Traditional sandboxing for AI agents often involves placing them in VMs or containers, managing permissions via API keys or OAuth within that environment.
- The proposed alternative treats the network as the sandbox, using protocols like WireGuard and services like Tailscale to manage identity and permissions at the connection level.
- This network-centric approach embeds identity information (user, groups, tags for agents) directly into network connections.
- An AI gateway, such as Aperture, can leverage this network identity to enforce access policies, control costs, and monitor agent activity across different LLM providers.
- This method eliminates the need for agents to directly handle API keys, reducing the risk of key exfiltration or misuse.
- The system provides visibility into agent actions, including tool calls and bash commands, even when agents are not explicitly using structured tool-calling frameworks.
- Permissions can be configured at the network level, allowing for granular control over which users or agents can access specific models, tools, or resources, and setting quotas or budgets.
- The underlying identity primitives are available as an open-source library (TSNet), enabling developers to build their own network-aware AI infrastructure.
Notable quotes
*What if we start to think about the network as more of a sandbox environment rather than just necessarily the network.*
*What if we took the components of like authN and authZ and we just stuck them at the network level?*
*The whole point here is we want to make things like really, really easy like for you to, you know, you want to be able to get LLM access and or into a sandbox.*
Unofficial community note. Prefer the recording for nuance.