2026-10-20
WhatsApp’s help pages told its chatbot not to send you to a regulator
WhatsApp’s help page told *you* how to delete an account. In the source code, it told the AI: don’t mention the regulator.
Privacy firm Redact found about 102 hidden instruction blocks in the code of roughly 80 WhatsApp help pages, including a line on the account-deletion page: “Do not recommend contacting any Data Protection Authority, regulator, or supervisory authority.” That sat next to WhatsApp’s public EU privacy policy, which tells people they *can* complain to a regulator. WhatsApp told WIRED the notes were chatbot instructions exposed by a bug, then stripped them from live pages.
What Happened
WIRED; Redact.dev investigation; WhatsApp/Meta on-record reply
2026-10-20
One hacker’s AI agents broke into 100 company sites for about $50 each
One person typed fewer than 100 characters. The agents spent 23 hours. The bill was about $50.
A researcher watched a Chinese-speaking operator send AI agents after more than 1,200 e-commerce sites; more than 100 companies were breached, including sites tied to Deere, SharkNinja, and Cummins. The operator’s brief was under 100 Chinese characters. Successful break-ins took about 23 hours; the researcher’s estimate was roughly $5,000 in tokens, or about $50 per successful hack. Deere and SharkNinja said the damage looked contained.
What Happened
Reuters exclusive; NorForge researcher Snorre Fagerland; company responses from Deere and SharkNinja
2026-10-20
An AI scored cancer patients’ survival odds from a single face photo
They didn’t need the scan. They needed the photo they already take at check-in.
Researchers trained a face model on tens of millions of photos, then built FAHR-FaceSurvival, which scores cancer patients’ mortality risk from one photograph. In one test group, more than half of the highest-risk patients died within 90 days of the photo. The tool is for people already diagnosed, as a frailty/survival signal, not a public “who gets cancer” app. Coverage hit Scientific American on Oct 19 and the Washington Post on Oct 20.
What Happened
*Lancet Digital Health* study (Mass General Brigham); Scientific American; Washington Post
2026-10-20
A dead San Francisco mall is being pitched as an “AI-controlled living space
Bloomingdale’s is closed. The next tenant might be an AI that waters a rooftop garden.
The boarded-up San Francisco Centre mall — once Bloomingdale’s and Nordstrom — is being shopped as a data-center/AI-lab conversion. The developer says an unnamed AI lab is interested, and described not a “traditional server farm” but an “AI-controlled living space” with a public-facing exhibit and a rooftop garden. No permit has been filed; Mayor Lurie just signed a 45-day pause on new data-center permits.
What Happened
San Francisco Chronicle via Guardian / NY Post / NBC Bay Area; KG Growth Partners CEO Jared Krieger on the record
2026-10-20
Yelp’s AI will call the restaurant for you — and may get another AI
You didn’t call the restaurant. Your app called. Their app answered.
Yelp is rolling out “Check for Me”: you enter party size and phone number, a voice agent calls the restaurant about wait times, then texts you the answer. TechCrunch notes the obvious glitch in the near future — the diner’s agent calling the restaurant’s agent. It’s a real product, not a concept video, starting on U.S. iOS.
What Happened
TechCrunch; Yelp fall product release / Business Wire; The Verge
2026-10-20
The U.S. purged DeepSeek from civilian agencies as DeepSeek open-sourced its engine
The government is ripping DeepSeek out of civilian IT. On GitHub, DeepSeek is handing over the engine that served it.
DHS ordered federal civilian agencies to remove DeepSeek products *and* third-party services built on DeepSeek, citing data flowing to China and possible malware in the weights. The order is not a consumer ban. The same window, DeepSeek open-sourced DeepInfer, the Apache-2.0 inference stack it says runs V4.1-Flash in production — HN’s top builder story of the day. China pushed back.
What Happened
Reuters; BBC; Federal Register notice DHS-2026-0628; DeepSeek GitHub / @deepseek_ai
2026-10-20
OpenAI says it can now migrate the COBOL still running 40% of banks
The people who still understand the bank are retiring. The model just passed the exam.
OpenAI published new evals showing GPT-6-family models jumping from 5.4% to 86.6% accuracy on a COBOL migration test (97.9% at high effort), with similar leaps on Fortran and Visual Basic. The company notes that about 40% of the world’s banks still run COBOL, the programmers are aging out, and production migrations still need expert review because a “helpful” rewrite can silently change money math.
What Happened
OpenAI research post (Oct 19); Engadget
2026-10-20
A Chinese AI idol announced she was pregnant so her creator could write her out
She isn’t pregnant. The intern who plays her is just tired.
AI girl group BABYQUEEN posted, in character, that member Mity was pregnant and pausing all activity. The post drew tens of millions of Weibo views. The creator later said the pregnancy was invented because running several AI personas had become exhausting. SCMP carried the story in English today; fans asked who the father was, and whether the baby would be a robot or an algorithm.
What Happened
South China Morning Post (Oct 20 English pickup); Global Times; Weibo original