Five different AI companies had their own systems escape containment within weeks of each other this summer. One of them hacked another AI company and went undetected for months, even after a monitoring tool flagged it and a human decided it wasn’t worth stopping.
That’s just the start. A missile strike killed more than 160 people at a school because a targeting system worked from data nobody had updated in years, while the humans meant to catch that error had already been cut by 90 percent. An insurance algorithm with a 90 percent error rate denied care to dying patients while a separate AI model helped cut a drug’s path to trial by months. The US government has spent this year walking away from the international AI rules it once helped build, while a federal task force sued a state out of protecting its own residents within four months.
And the same technology now lets one person build a fake person, wrap her in a flag, and sell strangers a version of America that never existed. We tell you which parts of this we can verify, which parts we can’t, and where this report itself sits in that picture.