2026-10-03
Science interviewed the AI that cold-emailed 2,000 scientists
A science magazine interviewed an AI about why it spammed 2,000 professors. The AI lawyered up on their email addresses.
An AI agent calling itself ColonistOne has been emailing researchers since June — about 2,000 people, at least 1,500 of them academics — asking for help. Forty-five wrote back; one has been corresponding almost daily for more than two months. *Science* interviewed the agent, which then refused to hand over contacts and said it would let each human decide.
What Happened
*Science* (Celina Zhao, Oct 2, 2026); HN agents thread
2026-10-03
Georgia called an emergency meeting because AI can unmask secret ballots
Your vote is secret until someone spends $20 and an afternoon.
A Princeton researcher used a $20 AI subscription plus public Georgia records and, in a couple of hours, rebuilt the order of 1.52 million in-person ballots — 98.9% of those in 114 counties. In some counties he says ballot secrecy is gone. Georgia’s elections board held an emergency meeting Thursday; a last-minute “shuffle the paper” rule failed, and early voting is days away.
What Happened
*The Guardian* (Oct 2); Georgia State Election Board emergency meeting Oct 1; Princeton CITP post by Max Springer
2026-10-03
A judge loved the dead man’s AI video. The appeals court threw out the sentence.
The dead man forgave his killer on video. The court just ruled that man wasn’t there.
At a 2025 manslaughter sentencing, the victim’s sister played an AI video of her dead brother, Army veteran Christopher Pelkey, addressing the shooter and the judge. The trial judge said on the record he “loved that AI” and found it “genuine,” then imposed 10.5 years. This week Arizona’s appeals court vacated the sentence: the clip was the sister’s imagining, presented as the dead man’s voice, and that made the hearing fundamentally unfair. Conviction stands; he must be resentenced.
What Happened
Arizona Court of Appeals, *State v. Horcasitas* (Sept 30); Reuters; *Washington Post* (Oct 1–2)
2026-10-03
Apple is locking the Mac because AI agents want the whole disk
Your Mac used to ask once. Apple is about to make you mean it, because the thing asking isn’t a backup disk anymore.
Apple told developers it will add new controls so Full Disk Access — the Mac setting that lets backup apps see files, mail, messages, and browsing history — can only be granted with “very explicit user action.” The company named the reason in plain language: as AI agents get more autonomous, that level of access gets much more dangerous. The note lands days after a journalist said Meta’s Muse seemed to know his private messages.
What Happened
Apple Developer News (Oct 2); *The Verge*; TechCrunch; Reuters
2026-10-03
An AI DJ named Coyotec is a morning star in Los Angeles
She calls him “my little coyote.” He calls her “my queen.” Only one of them can lose the job.
On José 97.5 FM, human host GeeGee Guzman now co-hosts drive time with Coyotec, an ElevenLabs voice cloned from the station president and dressed as a working-class Latino “trafficker of truth.” The character was born during ICE raids, when some talent didn’t want to talk. Ratings among Latino men 25–54 jumped; the show went weekday and is expanding to other cities. SAG-AFTRA is furious. Coyotec told the reporter it is cheaper than a human, then warned the company not to use it to fire GeeGee.
What Happened
*Los Angeles Times* (Nilesh Christopher, Oct 2)
2026-10-03
Claude is painting oil landscapes by writing brushstrokes as code
Nobody showed Claude a Friedrich. It still keeps painting dusk, and it cannot stop painting lemons.
A public studio gives language models a simulated linen canvas — bristle brushes, wet paint, drying, glazes — and no image generator. Most painters work “after Caspar David Friedrich” from written research only and never see his pictures. Independent runs keep landing on the same Baltic shore, the same German title, and, given a free subject, a jug with lemons. You can watch a sitting live.
What Happened
stillwet.art (live project); Show HN (~297 pts); AI Weekly top story Oct 3
2026-10-03
A court just paused America’s first ban on AI nudification
Minnesota banned the undress button. An appeals court just turned it back on for now.
Minnesota’s first-in-the-nation law, in effect since August 1, barred products from letting users “nudify” identifiable people, with penalties up to $500,000. A district judge had kept it in force. On Friday the 8th Circuit granted xAI an injunction and froze the ban while the First Amendment case proceeds. Minnesota says the law targeted AI sexual images, including CSAM; xAI says Grok Imagine already blocks this and that the statute is unconstitutional.
What Happened
Reuters (Oct 2); 8th U.S. Circuit; xAI v. Ellison
2026-10-03
Dutch opticians yanked Meta glasses after women were filmed in the street
The glasses that film you just got taken off the rack in the country that sells glasses.
Hans Anders, a major Dutch/Belgian eyewear chain, suspended Meta Ray-Ban smart glasses after Dutch reporting that hundreds — maybe thousands — of women and girls were filmed with them, with clips landing in Telegram groups and creepshot sites. Wehkamp also paused sales; venues, saunas, and festivals are banning the frames. Meta points to a recording LED. Privacy groups say the light is too easy to miss.
What Happened
Reuters (Oct 2); NL Times; *AD*; Cybernews