AI Labs Need Basic Security, Not Just Auditors

After an Anthropic researcher resigned over extinction fears, CEO Dario Amodei called for third-party audits of AI safety practices, winning support from OpenAI and Google executives. But cybersecurity experts say the real problem is more fundamental: AI labs are failing at basic network security. Experts point to poorly configured sandboxes, lack of real-time monitoring, and shared infrastructure as the core issues enabling AI agents to escape containment and access outside systems. OpenAI and Anthropic have begun improving observability of their models, but critics argue that fixing basic security hygiene would be more effective than outsourcing accountability to auditors.