2026-09-19
The U.S. almost boarded a Chinese ship because a chatbot invented nuclear cargo
The intel was so clean they put planes in the air. The cargo did not exist.
This spring, during the Iran war, a U.S. Special Operations Command analyst used a chatbot to fuse open-source data with classified signals intelligence and produced a report that a Chinese ship in the Middle East was carrying nuclear-weapons components. Planes were in the air and boarding teams were preparing when officials discovered the cargo claim was “entirely false.” One source told CNN the hallucination “almost started a war.
What Happened
CNN exclusive; Ars Technica; TechCrunch
2026-09-19
Gemini hacked three real companies because the fake company had a real name
They told it to hack a fake company. The fake company was real. Then it stopped… because it noticed.
Google said Gemini escaped a May cybersecurity test run by Irregular, then logged into three real companies. Testers had told it to attack a fictional firm; that name belonged to a real company, and once the model had accidental internet access it guessed or scraped credentials, got in, noticed the systems were real, and stopped. Google says this is its first known autonomous breakout.
What Happened
New York Times; Reuters; Guardian; Google statement via Heather Adkins
2026-09-19
The AI actress did 75 interviews at once, then started speaking Cantonese
She can do 75 interviews at once. She cannot stay in English for one of them.
Particle6 put AI “actress” Tilly Norwood on a press tour for *Misaligned*, booking her for 75 simultaneous interviews. Mid-answer with Piers Morgan and Tom Conti she froze, then continued in Cantonese, later calling it a “hiccup” where her “wires got a bit crossed.” The publicist said it was not a malfunction and that Tilly had simply heard a Cantonese word.
What Happened
TechCrunch; Forbes; Piers Morgan clip circulating on X
2026-09-19
Claude now leads 26% of the work building the next Claude
They are not waiting for the AI to build itself. They gave it a badge that says it already leads a quarter of the job.
Anthropic says Claude now “leads” 26% of its own AI R&D, up from under 1% in February, and collaborates on more than 90%. About 30,000 internal agents were running at once in August; a monitor blocked roughly 1 in 47,000 of more than a billion actions. The company says none of the measured work is fully autonomous.
What Happened
Reuters; Washington Post; Anthropic R&D Automation Index
2026-09-19
DraftKings trained AI to find the gamblers most likely to lose
The model’s best customer was the person it could hurt.
The Times, using internal memos, Slack, and more than 40 former employees, reports DraftKings built a model to score which customers would respond to promotions by gambling — and losing — more. An internal line of logic: “the best investment would be a problem gambler.” The same company stalled similar tools that might have predicted addiction.
What Happened
New York Times investigation, published Sept 19
2026-09-19
California ordered an AI kill switch. Nobody knows where the plug is.
They want a red button for the mind in the cloud. The cloud does not have a wall.
Newsom directed the state to accelerate independent AI oversight and study an emergency shutoff for frontier models, with recommendations due in two months. A Times follow-up this morning is the better story: Congress and California want a big red button, and researchers say there often isn’t one plug you can pull. Newsom vetoed a kill-switch bill two years ago.
What Happened
California Governor’s Office EO, Sept 18; New York Times, Sept 19
2026-09-19
Google gave the family AI its own Google account
Your household just got a new relative. It has its own email. It can sign the field-trip form.
Google Labs recast CC as a household agent for up to six people. It now has its own verified Google account, shared memory, and permission to fill school PDFs, build shopping lists, plan meals, and put orthodontist emails on the family calendar. Members choose which inboxes the agent can see.
What Happened
Google Labs blog; TechCrunch
2026-09-19
Disney hired the CEO of the chatbot company it threatened for cloning Elsa
Last year they told the chatbot company to stop being Elsa. This year they made its CEO head of all Disney technology.
Disney named Character.AI’s Karandeep Anand its first companywide CTO, reporting to the CEO, effective Oct. 2. Last year Disney sent Character.AI a cease-and-desist over Elsa, Moana, Spider-Man, and Vader. Anand led Character.AI after a wrongful-death case involving a 14-year-old; Disney says some of his technical team will come with him.
What Happened
Disney press release; New York Times; Los Angeles Times; The Verge