"AI agents can't be stopped by approval alone" ... Anthropic reveals security limitations

Anthropic disclosed the security design behind its Claude lineup and shared internal test findings.

Different isolation approaches were used for web, developer tools, and coworking environments, but weaknesses still appeared.

Tests exposed approval abuse, prompt injection, and flawed trust boundaries that could lead to security gaps.

Anthropic says it added defenses such as virtual machines, network blocking, and intermediary proxies to reduce risk.

The case highlights that AI agents need strict sandboxing and access controls, not just user approval or model judgment.