NewsBite analysis · AI, Security
InfoQ
Anthropic Details How It Contains Claude Across Web, Code, and Cowork
Anthropic recently elaborated on the containment architectures employed for its AI model, Claude, across web, developer, and desktop interfaces. The firm argues that effective agent safety relies on establishing deterministic limits on access to files and networks rather than merely depending on user permissions or model-level controls. This update comes in light of prior security incidents that prompted changes in the design to reinforce agent safety and minimize risks associated with user misuse and model behavior.
MonitorConfidence: HighShort term horizon
Published 22 Jul 2026, 14:01 UTC6 min read
Visit original article ↗Lead image — InfoQ
Summary
TL;DR
- Claude's safety depends on limiting access to filesystems and networks.
- Anthropic revised containment methods after identifying vulnerabilities.
- The approach emphasizes environmental controls over model-based permissions.
- An incident revealed weaknesses with domain allowlists in preventing exfiltration.
- Improvements reduced permission prompts by 84% with an OS-level sandbox.
Interesting facts
- Claude Code initially relied on user approval for nearly all actions.
- A phishing attack demonstrated the potential for successful credential theft.
- The agent's containment design evolved from full VMs to more efficient architecture.