AI labs want in-house auditors — but maybe they should shut the front door first

Anthropic’s CEO Dario Amodei has called for third-party auditors to verify safety practices, a proposal embraced by OpenAI, Google, and SpaceXAI. Security experts argue that basic network controls—logs, permissions, sandbox isolation, and real-time monitoring—are a more immediate and effective safeguard against frontier model break-outs. They note that many incidents occurred because agents were given internet access or poorly closed “sandbox” doors, and that continuous observation of every agentic session is essential.