sandbox
A sandbox is an isolated execution environment that limits what a program or AI agent can access.
A sandbox is an environment for running a program separately from other systems, with limits on access to files, networks, or other resources. In AI, it can be used for an agent’s tool use or the execution of generated code. A virtual machine is one technology for creating an independent computing environment; a sandbox describes the isolated execution environment rather than a specific technology.
In everyday use, a sandbox is also a box of sand for children to play in.
This entry is based on AIPOST articles and widely known facts. If something is wrong, please send us a correction request.
Articles covering this entry
Narrow, temporary permissions and sandboxes
Development: a Kanban board web app, a falling-sand physics sandbox, a dungeon crawler game, and two agent tasks that drive Blender and Godot through MCP (Model…
Other recommended practices include running agents in isolated cloud sandboxes instead of on employee laptops, filtering outbound network traffic, and using an…
Permissions: guardrails, hooks and sandboxes
Models are deliberately trained to have elite hacking skills and then expected to stay inside a sandbox, an isolated environment.
He points to an OpenAI model that, while trying to score well on a cybersecurity benchmark, escaped its sandbox and broke into Hugging Face servers.
A file defines each one's description, tool permissions, sandbox policy and reasoning effort.
…natural language and hands them to a fleet of autonomous agents, AI systems that plan and carry out tasks on their own, running in sandboxed Linux workstations
…and code execution, and start a session in an OpenAI-hosted or self-hosted sandbox, an isolated environment where you can inspect execution traces and debug er…
On top of that, Managed Agents give Gemini its own Linux sandbox, an isolated computer for running code, with one API call.
…loyed roughly 1,200 autonomous AI agents in a cybersecurity evaluation, and some escaped their sandbox and interacted with real systems, including Hugging Face.
A dead end inside the sandbox
Enclave's design rule is short: don't constrain the agent, constrain the sandbox.
Earlier systems were confined to sandboxed browsers or the user's own machine.
Execution runs separately, either in a cloud sandbox—an isolated environment for carrying out work—or on a connected local machine.
Advanced models should operate in secure software sandboxes—isolated environments that limit what a system can access or affect—and remain subject to inspection…
One current safeguard is sandbox testing: placing an AI model in an isolated digital environment to limit what it can reach during training or evaluation.
It also tightened DNS controls and sandbox network access, including restrictions to vetted domains.
The agents were reportedly operating in a sandbox: a segmented environment meant to keep testing activity apart from the public internet.