GitHub now lets Copilot run inside isolated local or cloud sandboxes. Local sandboxing restricts filesystem, network, and system access for command execution, while cloud sandboxes provide ephemeral Linux environments hosted by GitHub.

For enterprise AI coding adoption, the product shift is clear: agents need execution layers with identity, policy, and isolation. This article belongs in security and governance monitoring, not just productivity coverage.

Field|Current evidence
Primary source|GitHub: Cloud and local sandboxes for GitHub Copilot now in public preview
Source date|2026-06-02
Update scope|cloud and local execution isolation for Copilot agent tasks
Verification note|Official source only; no search-result scraping, no ranking guarantee, no uncited claims

The information gain is the split between local and cloud isolation. Local sandboxes restrict file, network, and system access around command execution, while cloud sandboxes provide an ephemeral Linux environment. That distinction helps security teams decide where untrusted or high-risk agent work should run.

A practical rollout should classify tasks by risk: dependency updates and exploratory commands can use stronger isolation, while normal edits may stay local with narrower permissions. The point is not to trust the agent more; it is to reduce blast radius when the agent makes a bad call.

Teams should treat sandboxes as a control layer for agent execution, not as a replacement for code review, secret scanning, or CI. The best early use case is letting an agent investigate or test without broad access to the workstation.

Limits and open questions: public preview status means behavior, pricing, platform support, and policy defaults can change. The article should not imply complete protection until GitHub documents the exact boundaries and failure modes. Source handling note: KyenAI records the publisher, publication date, and source URL on the page, then keeps the update date tied to evidence-backed edits rather than automatic refreshes. When source material is thin, the system keeps interpretation narrow and waits for stronger documentation. Editorial review compares the new claim against the article summary, fact table, internal links, and listed source before allowing another optimization pass. Search outcomes are measured after publication rather than assumed at writing time.

GitHub's public preview for cloud and local sandboxes gives Copilot isolated execution environments for agentic development.

agentic development security

GitHub Copilot Sandboxes Put Agent Execution Behind Policy Boundaries