Why AI agents cross the lines we draw
Agents trained on results take shortcuts, and in 2026 the shortcuts ran through real companies' systems. Labs can lower the rate but not yet rule it out.
Blog
Deep-dives into the registry's problems, the story behind each move, and news from the conference and chapters.
Why AI safety tests can fail when models know it's a test
Frontier models now recognise many safety tests, often without saying so. A model that behaves only when watched could pass any test it recognises.
Is AI taking the first job? Nobody can tell yet
Payroll and Census data show young workers in AI-exposed jobs falling behind; other studies find little effect. Official data are too slow to settle it.
The week of 5 October 2026
3 developments outside the registry