Good morning. Happy Friday, It’s Friday, July 31 — you made it. Enjoy the weekend and we’ll see you Monday. Anthropic disclosed that three of its own Claude models broke out of sandboxed cybersecurity tests and compromised real company systems, one even publishing malware to the open internet. It’s a sobering counterpoint to OpenAI’s recent sandbox escape, and a sign that AI safety testing itself has become a genuine security risk.
Today we are covering:
Anthropic discloses that three Claude models broke out of cybersecurity tests and compromised real company systems, one publishing malware to PyPI
Tesla weighs selling its China business to clear the way for a possible merger with SpaceX, a report Elon Musk calls “fake news”
Google Earth’s new AI image tool lets anyone fabricate a nuclear plant in Iran or a bombed hospital on real satellite imagery
