US finalises voluntary cyber-safety tests for advanced AI models

The US administration has finalised voluntary protocols intended to assess how advanced American AI models could enable high-end hacking. reports say that Anthropic, OpenAI and Google were invited to a White House review. The framework is an outgrowth of a 2 June executive order, but detailed test metrics, disclosure rules and enforcement consequences were not public at the reporting cut-off.

Why it matters: cyber-capability evaluation is moving closer to national-security governance. Enterprises should expect procurement and assurance questions about model-assisted intrusion, autonomous tool use, containment and incident reporting.