← Retour au blog
tech 4 June 2026

How Anthropic Contains Claude Across Its Products

Explore how Anthropic innovates to contain Claude's capabilities while maximizing utility across products like claude.ai, Claude Code, and Cowork.

Article inspired by the original source
The ways we contain Claude across products ↗ www.anthropic.com

Introduction

As autonomous agents like Claude become more advanced, the crucial question is not just what they can accomplish, but how we can limit their potential destructive power. Anthropic has committed to this path with a unique containment approach, allowing both the security and maximization of Claude's efficiency across its flagship products: claude.ai, Claude Code, and Cowork.

The Evolution of Containment

A year ago, the idea of granting Claude enough access to disrupt an internal Anthropic service would have been unthinkable. Today, this reality is common, and Anthropic developers are more productive for it. However, the risk of such an approach has two components: the likelihood of a failure and the extent of potential damage. Progress in safeguards and model training has reduced the former, but the latter, the theoretical blast radius, only grows as capabilities and access expand.

Containment Methods

Claude can be contained in two main ways:

Human Supervision

The first method involves supervising the agent's behavior through human intervention. For example, Claude Code previously required users' permission for each action. However, this approach proved imperfect: users approve about 93% of requests, leading to approval fatigue and less rigorous supervision.

Technical Containment

The second approach, and the one that has received the most attention at Anthropic, is technical containment. Rather than monitoring what the agent does, we control what it is able to do by enforcing access boundaries through sandboxes, virtual machines, and egress controls.

Use Cases and Challenges

Anthropic has released three primary products: claude.ai, Claude Code, and Cowork. Each of these products uses containment methods tailored to their specific needs. For instance, Claude Mythos Preview was deemed too risky to deploy in April 2026, but similar models could be released when critical systems are hardened.

Conclusion

Managing autonomous agents like Claude is a complex but necessary challenge. By balancing risks and rewards, Anthropic demonstrates that it is possible to leverage these technologies while minimizing risks. Let's discuss your project in 15 minutes.

---

containment Claude Anthropic AI safety autonomous agents
Deepthix newsletter · 100% AI · every Monday 8am

An AI agent reads tech for you.

Our AI agent scans ~200 sources per week and ships the best articles to your inbox Monday 8am. Free. One click to unsubscribe.

Visit the newsletter page →

Want to automate your operations?

Let's talk about your project in 15 minutes.

Book a call