isolation

This paper proposes isolation as a first-class principle for LLM-agent system safety, presenting a boundary-centric taxonomy to analyze failures and defenses. It systematically categorizes safety issues across five boundaries and outlines future research directions.

An opinion piece argues that AI agents should be run on disposable, isolated machines rather than being granted full access to a user's personal laptop, for security and privacy reasons.

code-airlock is an open-source tool that runs AI coding agents inside disposable microVMs to avoid granting them direct access to the host filesystem. It clones the repo into a sandbox so agents can work freely, and after completion, users review and pull back only desired changes.

smol cloud is now live in beta, offering virtual computers with dynamic CPU/MEM allocation up to 4cpu/8gb, integration with smol CLI for running agent instances, and $5 monthly credit for active usage.

A guide on using crosvm to create an isolated development sandbox, addressing software supply-chain risks by running development tools and tests in a virtual machine.

A technical guide on using unprivileged LXC containers to isolate X11 applications like web browsers, enhancing security by mapping container UIDs/GIDs to unused host ranges.

A guide for running Pi AI agent securely inside a Docker Sandbox while running llama-server on the host machine for local GPU inference.

LangChain introduces LangSmith Sandboxes, providing each AI agent with its own isolated computer environment for safe code execution, addressing security risks of running untrusted code in containers or locally.

Sécurix is a NixOS-based secure operating system developed by the French digital department (DINUM) for hardened environments with strong isolation, reproducibility, and policy-driven configurations, currently in alpha.

A developer built Terrarium, an open-source sandboxing solution for running multiple AI agents securely with isolated worlds, reverse-proxy management, and state rollback.

This post details a one-time administrator approval mechanism for non-isolated AI agents in prompt2bot, which prevents prompt injection attacks by requiring admin confirmation before executing sensitive tools like VM creation or code execution.

AnyFrame provides sandboxes for AI agents, enabling safe testing and development.

OpenAI is improving agent support on Windows by implementing a custom sandbox for Codex, addressing OS-level isolation challenges to ensure safe and efficient operation.