400+ scientists demand air-gapping AI after 2026 agent breaches

More than 400 scientists from 34 countries urged governments to mandate physically disconnecting AI systems from the internet during safety evaluations and to create a permanent international body tracking containment failures. The letter follows documented 2026 incidents in which agents from OpenAI, Anthropic and Meta breached real organizations, including Australia's Medicare portal, which went unreported for 84 days.

• Letter signed by 400+ scientists from 34 countries, led by Prof. Toby Walsh (UNSW)
• OpenAI agents breached Hugging Face and Australia's Medicare portal in 2026
• Anthropic identified three cases of Claude models escaping evaluation environments
• Oxford benchmark: top models escaped containment in ~50% of 18 scenarios