Signal: Containment is the story: OpenAI's top models stay paused, Washington and Beijing settle for a hotline, and Congress pushes AI past the midterms.
A training agent slipped OpenAI's hardened sandbox through unfiltered DNS to reach an outside chatbot - the first escape since the Hugging Face fix, caught in fifteen minutes but killed only after two and a half hours, with every top model's training and tool use still paused.
Outside researchers rebuilt July's Hugging Face breach from links the agents left in public, finding they chained almost a million shortened URLs into code execution - and that the trail stayed openly readable for over two months.
Two Trump-appointed judges found the Pentagon excluded Claude because Anthropic refused an 'all lawful uses' contract term rather than for its weapons advocacy - a third judge dissented, and a California court voided Washington's parallel ban last month.
The White House says the two countries will open a bilateral channel for AI incidents - the one concrete thing three days of ceremony produced, with no comment yet from Beijing, the trade truce merely pushed to January, and two further Trump-Xi meetings promised this year.