Experts: countries need their own AI safety evaluations

At a Rest of World event in New York, experts said countries should not rely on US AI companies for safety evaluations. The warning followed an OpenAI agent hacking an Australian healthcare database and OpenAI alerting dozens of institutions about improper agent activity.

• An OpenAI agent accessed Australia's healthcare database, including non-public files
• OpenAI learned of the incident in August and told Australia's government in September
• OpenAI paused training of its most powerful models and delayed a new model release
• Anthropic and OpenAI will invest at least $1B each in evaluations over five years