Anthropic is shifting from model-level safeguards toward hard environmental controls to secure its AI agents. By implementing filesystem isolation and rigorous network restrictions, the company aims to mitigate risks that cannot be reliably managed by human approval prompts or internal model classifiers.