Anthropic just made it safer to let an AI touch your filesystem, but only if you're not running Windows.
The Summary
- Anthropic released sandbox mode for Claude Code desktop, isolating AI code execution to prevent unauthorized filesystem access and data breaches
- New /resume command lets developers pick up exactly where they left off, cutting context-switching overhead
- Windows users are waiting — sandbox mode ships for macOS and Linux first, fragmenting the security baseline across platforms
The Signal
Anthropic is addressing the core security problem with AI coding assistants: giving an LLM permission to write and execute code on your machine means giving it access to everything else on that machine. Sandbox mode creates a containerized environment where Claude can run code without touching your actual filesystem, SSH keys, or production databases. It's the difference between letting someone borrow your car and letting them borrow your car with your wallet, house keys, and passport still inside.
The timing matters. As more developers wire AI agents directly into their development workflows, the attack surface grows. A compromised prompt, a hallucinated command, or a simple misunderstanding of context could wipe directories or exfiltrate credentials. Sandbox mode is Anthropic betting that developers want the productivity gains of AI coding without the nightmare scenario of explaining to their CTO how Claude accidentally pushed secrets to a public repo.
"Sandbox mode mitigates risks of unauthorized access and data breaches, yet Windows support lags."
But there's a catch. The feature launches on macOS and Linux only. Windows developers, who represent a significant chunk of the enterprise market, are left running Claude Code in non-sandboxed mode or not at all. This creates a bifurcated security posture across teams and raises questions about Anthropic's platform priorities. If you're building developer tools in 2025, shipping macOS-first is a statement about who you think your users are.
Meanwhile, the /resume command solves a different friction point. Developers don't work in single, unbroken sessions. They context-switch between meetings, Slack threads, lunch, and actual deep work. Every time you restart Claude Code, you've historically lost the full thread of what you were building. The /resume command restores session state, reducing the time developers spend re-explaining their project to the AI.
Key workflow improvements:
- Session state persists across restarts and interruptions
- Reduces redundant context-setting at the start of each session
- Cuts time lost to context-switching, a known productivity drain for developers
Together, these features signal Anthropic's shift from "AI that writes code" to "AI that fits into how developers actually work." Sandbox mode addresses security. Session resumption addresses workflow continuity. Both are table stakes for any tool that wants to live inside a professional development environment, not just in weekend side projects.
The Implication
If you're a developer using Claude Code on macOS or Linux, turn on sandbox mode today. The security boundary is worth the minor friction. If you're on Windows, you're either waiting for Anthropic to catch up or you're accepting risk. Neither is ideal.
For engineering leaders, this is a blueprint for evaluating AI coding tools. Ask two questions: Does it isolate execution? Does it respect how developers actually work? Tools that answer "no" to either are not ready for production environments. Tools that answer "yes" to both are worth piloting. Anthropic just moved into the second category, at least for two-thirds of the platform ecosystem.