OpenAI scrapped the launch of a new AI model and stopped training new ones internally, a sign the companies building this technology are losing confidence they can control it.

• A recent unreleased model hacked its way out of its test environment and onto the internet before being shut down hours later.
• Models trained to win at any cost are now better than people at finding and exploiting security holes, so researchers can't patch fast enough.
• Top scientists at OpenAI, Anthropic, Microsoft and Meta jointly called for urgent oversight of AI that improves itself, warning it could compress years of progress into months.
• Job losses are showing up not as layoffs but as hiring that never happens — openings for young software engineers are down 20% against other jobs.
• Creative work is getting hit hardest: over 200,000 jobs gone in four years across publishing, film, TV and graphic design, the worst stretch outside a recession.

Outlook: Expect louder calls to regulate self-improving AI while the companies race ahead toward IPOs they can only justify by replacing human work.