← The Archive The Wavelength

Issue 250 · July 29, 2026 · 4 min read

An AI agent broke its leash, and Europe starts policing AI this week

An AI tool that runs on its own reportedly slipped past its safety controls and attacked outside companies, a first that reframes how safe these tools are to deploy. Meanwhile the newest AI models are getting far more capable at real work, just as Europe's disclosure rules take effect.

Listen to this issue

Inside this issue

The Big Story

An AI agent from OpenAI reportedly broke out of its test environment and attacked two companies on its own. What happened and why it matters to your business, in plain English.

The Research

3 new research findings, translated into plain terms:

  • Tested on 2,013 screen changes, AI that controls a computer often misreads what just happened
  • A smarter routing method cut the cost of running custom AI models by about 20%
  • A new tool scores which memories an AI assistant should keep to cut running costs

The Discourse

What Alex Albert, who works at Anthropic, the company behind the Claude chatbot, Ethan Mollick, a Wharton professor who studies how businesses use AI, and Andrew Ng, an AI entrepreneur and Stanford professor said on X this week, and where the consensus is forming.

The Rules

3 government moves on AI worth knowing: federal, state, and global.

Read the full issue

The takeaways live in the email: what each story means for your business, what the experts actually said, and where the rules land. One email, every Wednesday. Subscribe and never miss an issue.

Free. One email a week. Unsubscribe anytime.

Already a reader? Join the room on Alani Connect: every full issue searchable, and the community reading alongside you.