August 2026 Summaries

Tailscale describes the “lethal trifecta” of AI-agent security risks: access to private data, exposure to untrusted content, and the ability to communicate externally, which together can enable prompt-injection attacks and data leaks. The post argues that common approaches either leave agents overly permissive, make them too restricted to be useful, or rely on frequent approval prompts that users eventually ignore. It proposes combining Aperture’s LLM and MCP gateway with Tailscale’s identity-based network and device posture controls to separate sensitive-data access from unrestricted internet access. Organizations can label connectors according to whether they contain customer data, assign sandboxes postures indicating whether they have internet egress, and use application capability grants so internet-connected agents can access only non-sensitive connectors while isolated agents may access sensitive data. By keeping control of LLM, API, and MCP access outside the agent harness, the design aims to make bypasses more difficult without requiring constant user intervention, although it is intended to defend against externally manipulated agents rather than malicious insiders or administrators.