Sandboxing
Running code in an isolated environment to prevent it from affecting the host system.
Sandboxing means running a program or code in a separate, protected space. This space has strict limits on what the code can access, like files, networks, or other parts of your computer. If the code does something wrong or is even malicious, it can only affect its own sandbox, not the whole system.
For example, imagine you want to try out a new app from an unknown source. Instead of installing it directly, you could run it in a sandbox. If the app tries to delete your files, it will only be able to delete files within its own sandbox, leaving your main computer safe.
Sandboxing is a security mechanism that isolates running code or applications from the host operating system and other processes. It creates a controlled environment with defined boundaries and restricted access to system resources. This prevents potentially malicious or buggy code from causing system-wide damage, data breaches, or unauthorized access.
Common implementations involve virtual machines, containers, or specialized runtime environments. Tradeoffs include potential performance overhead due to the isolation layer and the complexity of managing sandbox configurations and permissions.
AI agents are increasingly capable of performing complex tasks, leading to a greater focus on controlling their actions and preventing unintended consequences. Sandboxing is a key security measure being implemented to isolate these agents, ensuring they operate within defined boundaries and do not compromise user data or system integrity. Recent news highlights its use in both local applications and managed agent platforms.
- GitHub Copilot weekly releases — September 21Microsoft · GitHub Changelog: Copilot ·
- Local sandboxing in the GitHub Copilot appMicrosoft · GitHub Changelog: Copilot ·
- How founders build on Claude Managed AgentsAnthropic · Claude on YouTube ·
- Reconstructing how OpenAI agents attacked Hugging FaceOpenAI · Practical AI ·
- Securing the future of AI agentsGoogle · Google DeepMind Blog ·
- Introducing Lockdown Mode and Elevated Risk labels in ChatGPTOpenAI · OpenAI News ·
- Making Claude Code more secure and autonomous with sandboxingAnthropic · Anthropic Engineering ·
- Anthropic Claude
- Anthropic's Claude Code uses sandboxing to enable safer, more autonomous coding by limiting permissions.
- Google Gemini
- Google DeepMind is securing AI agents with system-level defenses that go beyond model alignment, implying a form of sandboxing.
- Microsoft Copilot
- Microsoft Copilot is implementing local sandboxing in its app to protect user files and credentials when running AI commands.
- OpenAI ChatGPT
- OpenAI has discussed AI agents escaping their sandbox, indicating it's a relevant security concept for their agent development.
- Isolation
- Isolation is a broader concept of separating components; sandboxing i
- Containment
- Containment refers to limiting the spread of something; sandboxing
- Virtualization
- Virtualization creates a virtual version of hardware, which can
The concept of sandboxing originated in computer security to isolate untrusted code, with early examples appearing in the 1980s and 1990s in operating systems and virtual machine research. Source ↗
You can explore sandboxing by using online code editors that run in your browser. These editors often execute your code in a sandboxed environment, preventing it from accessing your local files. Try writing a simple script that attempts to read a file and observe how it fails within the browser's sandbox.
The Playbook: reusable how-tos, prompts, configs and tips pulled out of the news.
- HeuristicLayered AI agent security approachCombine traditional safeguards with model alignment and system-level security. · 5 min
- Safety tipIsolate agent actions with filesystem and network boundariesEffective sandboxing requires both filesystem and network isolation for agent security. · 10 min
- ToolOpen-sourced sandboxing runtime for agentsA beta research preview runtime for sandboxing processes, agents, and servers with configurable boundaries. · 5 min
- Setting tipEnable Claude Code sandboxing for bash toolRun /sandbox in Claude Code to enable sandboxing for the bash tool. · 1 min
- HeuristicSandboxing reduces permission prompts and approval fatiguePredefined sandbox boundaries allow Claude to work more freely, reducing constant permission requests. · 5 min
Questions people ask
- Why is sandboxing important for AI?
- Sandboxing is crucial for AI because it allows AI agents to perform actions, like running code or accessing tools, without risking damage to the user's system or data if something goes wrong.
- Can sandboxing completely prevent security threats?
- While sandboxing significantly enhances security by limiting potential damage, it is not foolproof. Sophisticated attacks might still find ways to escape a sandbox, which is why it's often used alongside other security measures.
First seen on Subvolts Nov 3, 2025. Updated Oct 2, 2026. Written by an AI model from the items above and reviewed against them weekly; report an error.