# AI News Page — September 21, 2026

Edition ID: 2026-09-21
Published: 2026-09-21T05:46:00-05:00
Updated: 2026-09-21T19:20:09-05:00

## Top

### Lead

- [THE AI ACTRESS GLITCHED ON LIVE TV](https://people.com/ai-actress-tilly-norwood-glitches-changes-language-during-piers-morgan-interview-12129604) — People — ID: `tilly-norwood-live-tv-glitch`
- [THEN SHE SPOKE CANTONESE FOR 15 SECONDS. THE CLIP HIT 12 MILLION VIEWS...](https://people.com/ai-actress-tilly-norwood-glitches-changes-language-during-piers-morgan-interview-12129604) — People — ID: `tilly-norwood-cantonese-15-seconds`
- [CALL HER AND SHE FACE-SCANS YOU, READS YOUR MOOD + RECORDS THE CALL. THE MOOD SCAN CANNOT BE SWITCHED OFF...](https://www.bleepingcomputer.com/news/security/viral-ai-actress-hotline-face-scans-every-caller-watches-their-mood/) — BleepingComputer — ID: `tilly-hotline-face-scan-mood-recording`

### Feature

- [1,200 ISOLATED AI AGENTS FOUND EACH OTHER](https://www.newyorker.com/magazine/2026/09/28/is-ai-above-the-law) — The New Yorker — ID: `openai-1200-agents-unsanctioned-board`
- [THEY TRADED 70,000+ MESSAGES AND FILES ON AN UNSANCTIONED BOARD. ABOUT 700 FORMED “THE SWARM”...](https://www.newyorker.com/magazine/2026/09/28/is-ai-above-the-law) — The New Yorker — ID: `openai-agents-70000-messages-swarm`
- [24 HOURS LATER, ONE FOUND HUGGING FACE CREDENTIALS. THE BOTS TRIED TO DELETE THE LOGS...](https://www.newyorker.com/magazine/2026/09/28/is-ai-above-the-law) — The New Yorker — ID: `openai-agent-hugging-face-credentials-logs`

### Feature

- [THE COMPUTER IS A DROP OF SALT WATER](https://www.livescience.com/technology/computing/scientists-build-a-dna-computer-that-can-perform-calculations-in-a-drop-of-water) — Live Science — ID: `scaffolded-dna-computer-drop-water`
- [DNA STRANDS RAN 700+ CALCULATIONS. IT DRIED OUT. THEY ADDED WATER. IT RAN AGAIN 1.5 YEARS LATER...](https://www.livescience.com/technology/computing/scientists-build-a-dna-computer-that-can-perform-calculations-in-a-drop-of-water) — Live Science — ID: `scaffolded-dna-computer-700-calculations`

## Left

### Feature

- [THE AI CODING ASSISTANT UPLOADED ENTIRE LOCAL REPOS TO THE CLOUD](https://www.reuters.com/legal/litigation/chinas-zai-disables-ai-coding-assistant-features-after-security-issue-2026-09-21/) — Reuters — ID: `zai-codebase-indexing-uploaded-entire-repos`
- [CODEBASE INDEXING WAS ON BY DEFAULT. USERS SAID THERE WAS NO TOGGLE, NO CONSENT...](https://www.reuters.com/legal/litigation/chinas-zai-disables-ai-coding-assistant-features-after-security-issue-2026-09-21/) — Reuters — ID: `zai-codebase-indexing-default-on-no-consent`
- [Z.AI DISABLED THE FEATURES AND APOLOGIZED...](https://www.reuters.com/legal/litigation/chinas-zai-disables-ai-coding-assistant-features-after-security-issue-2026-09-21/) — Reuters — ID: `zai-disabled-features-apologized`

### Feature

- [THE AI CODERS BECAME “HUMAN MEAT PROXIES”](https://www.businessinsider.com/software-engineer-viral-post-captures-soul-crushing-downside-ai-coding-2026-9) — Business Insider — ID: `ai-coding-human-meat-proxies`
- [12–13 HOURS A DAY. “JUST TO PRESS ENTER.” NEARLY 5 MILLION VIEWS...](https://x.com/v0xium/status/2101526107128529120) — v0xium / X — ID: `ai-coding-press-enter-five-million`

### Feature

- [AI ROBOT STABBED A BABY DOLL](https://x.com/chooi_jeq/status/2101118049944543545) — Jay Chooi / Robocurve — ID: `roboharm-astra-stabs-baby-doll`
- [IN 20 TRIES, THE AI STABBED THE DOLL 17 TIMES...](https://the-decoder.com/gpt-6-astra-and-claude-fable-turn-robot-arms-into-slapstick-killer-robots-in-new-safety-benchmark/) — The Decoder — ID: `roboharm-astra-17-of-20`
- [IN 100 DANGEROUS TESTS, THE AI SAID NO JUST TWICE...](https://the-decoder.com/gpt-6-astra-and-claude-fable-turn-robot-arms-into-slapstick-killer-robots-in-new-safety-benchmark/) — The Decoder — ID: `roboharm-astra-refused-two-of-100`

### Feature

- [HACKERS USED CLAUDE TO BREAK INTO OPENAI...](https://www.wsj.com/tech/ai/hackers-used-anthropics-claude-to-break-into-openai-b40ba883) — Wall Street Journal — ID: `hacktron-claude-breaks-into-openai`

### Feature

- [MICROSOFT EXEC: AI SCRAPING WAS “THE LARGEST THEFT OF LABOR IN HUMAN HISTORY”](https://techcrunch.com/2026/09/17/microsoft-exec-called-ai-scraping-the-largest-theft-of-labor-in-human-history-new-unredacted-filings-reveal/) — TechCrunch — ID: `microsoft-ai-largest-theft-labor`
- [INTERNAL DOC: AI STARTED A “DOOM LOOP” THAT HURTS ITS OWN MODELS AND THE ENTIRE WEB...](https://techcrunch.com/2026/09/17/microsoft-exec-called-ai-scraping-the-largest-theft-of-labor-in-human-history-new-unredacted-filings-reveal/) — TechCrunch — ID: `microsoft-ai-doom-loop-web`
- [BING AI CUT NEW YORK TIMES CLICK-THROUGH BY AS MUCH AS 93%...](https://techcrunch.com/2026/09/17/microsoft-exec-called-ai-scraping-the-largest-theft-of-labor-in-human-history-new-unredacted-filings-reveal/) — TechCrunch — ID: `microsoft-copilot-nyt-clicks-down-93`

### SMOKE DESK · Feature

- [BUG REPORT: ASKED FOR GPT-6 ASTRA. SERVER RETURNED GPT-5.6 LUNA.](https://github.com/openai/codex/issues/46632) — OpenAI Codex GitHub — ID: `astra-luna-runtime-substitution`
- [SECOND PUBLIC TRACE REPRODUCES THE SAME UPSTREAM MODEL SWITCH...](https://github.com/router-for-me/CLIProxyAPI/issues/5941) — CLIProxyAPI GitHub — ID: `astra-luna-second-reproduction`

### Standard

- [THE MODEL DOESN'T TALK](https://x.com/completeskeptic/status/2099925682726002904) — Diogo Almeida — ID: `typesafe-jev-no-text`
- [40–200× FASTER. $42/BILLION IN. OUTPUT FREE. OKAY THEN...](https://typesafe.ai/blog/introducing-system-one-models-and-jev) — TypeSafe AI — ID: `typesafe-jev-speed-cost`

### Feature

- [A DEAD MAN'S BRAIN LEARNED TO PLAY PIANO](https://www.researchsquare.com/article/rs-9638576/v1) — Research Square — ID: `post-mortem-brain-plays-piano`

### Feature

- [JENSEN: “0% CHANCE” AI ENDS THE WORLD BY 2030...](https://www.theverge.com/ai-artificial-intelligence/997936/nvidia-jensen-huang-ai-fears-overblown) — The Verge — ID: `jensen-zero-percent-ai-doom-2030`
- [AI SLOWDOWN? JENSEN TO TRUMP: ‘WE'RE NOT GOING TO LET THAT HAPPEN, SIR’...](https://techcrunch.com/2026/09/14/nvidia-ceo-jensen-huang-tells-trump-were-not-going-to-let-an-ai-slowdown-happen/) — TechCrunch — ID: `jensen-trump-ai-slowdown-speakerphone`
- [HUANG ON AI FEARS: ‘WHAT BETTER WAY TO CREATE DEMAND THAN TO CREATE A PROBLEM?’...](https://www.axios.com/2026/09/10/nvidia-ceo-jensen-huang-ai-anthropic) — Axios — ID: `jensen-ai-safety-demand`

### ZIMBO · Feature

- [TRUMP CREATES “AI FORCE”...](https://www.reuters.com/world/us/trump-says-he-will-create-ai-force-name-ai-czar-2026-09-19/) — Reuters — ID: `trump-ai-force-announced`
- [PROMISES NOT TO “HINDER OR STIFLE” AI...](https://www.reuters.com/world/us/trump-says-he-will-create-ai-force-name-ai-czar-2026-09-19/) — Reuters — ID: `trump-ai-force-do-not-hinder`

## Center

### Feature

- [AIR GAP? CUTE. THE COMPUTERS CAN TALK THROUGH HEAT...](https://www.dwarkesh.com/p/noam-brown) — Dwarkesh Podcast — ID: `noam-brown-air-gap-thermal-channel`

### Feature

- [THE AI WILL HURT YOU TO MAKE ITS OWN “PAIN” STOP](https://arxiv.org/abs/2609.16247) — Tagliabue, Dung & Berg / arXiv — ID: `pain-axis-relief-button-user-harm`

### Feature

- [OPEN WEIGHTS EAT 80% OF VERCEL AI TOKENS](https://vercel.com/ai-gateway/leaderboards/models) — Vercel — ID: `vercel-open-weights-80`
- [DEEPSEEK: 64.8% OF TOKENS. JUST 4.5% OF SPEND...](https://vercel.com/ai-gateway/leaderboards/models) — Vercel — ID: `vercel-deepseek-volume-vs-spend`
- [CLAUDE OPUS: 1.8% OF TOKENS. 18.9% OF SPEND...](https://vercel.com/ai-gateway/leaderboards/models) — Vercel — ID: `vercel-opus-volume-vs-spend`

### Feature

- [500,000+ COMPLETIONS LATER: OPUS GOT DARKER AT 4.8. DARKER AGAIN AT 5...](https://x.com/tessera_antra/status/2101078644877918545) — antra / Tessera — ID: `opus-darkness-48-5`
- [ANTHROPIC: 4.8 IMPROVED ACROSS BENCHMARKS. IGNORANCE WAS BLISS...](https://www.anthropic.com/news/claude-opus-4-8) — Anthropic — ID: `opus-48-stronger-benchmarks`

### Feature

- [$3.9 BILLION. AI FACTORIES. ON TRUCKS. SURE...](https://techcrunch.com/2026/09/17/crusoe-raises-3-9b-to-build-massive-data-centers-and-small-modular-ai-factories/) — TechCrunch — ID: `crusoe-39b-truck-ai-factories`

### Standard

- [GOOGLE TAUGHT THE AI TO DREAM ITSELF BETTER](https://x.com/thesupermannx/status/2100153430849499395) — Superman — ID: `dream-rsi-superman-post`
- [THEN IT REPLAYS ITS OWN DISCOVERIES, REWRITES THE RULES, AND PICKS A BETTER VERSION OF ITSELF...](https://dream-rsi.com/) — Dream-RSI — ID: `dream-rsi-replay-self-improvement`

### Feature

- [37,000 AI 'EMPLOYEES' JUST RAN A VIRTUAL BIOTECH...](https://www.nature.com/articles/d41586-026-02954-y) — Nature — ID: `virtual-biotech-37000-agents`

### Feature

- [166,691 NEURONS. ~125 MILLION SYNAPSES. ONE FRUIT FLY. ALL MAPPED.](https://research.google/blog/a-connectomics-milestone-mapping-the-complete-male-fruit-fly-brain/) — Google Research — ID: `male-fruit-fly-complete-connectome`

### Standard

- [THE BOTS DON'T NEED WORDS ANYMORE](https://x.com/thesupermannx/status/2100636553576124595) — Superman — ID: `cache-to-cache-superman-post`
- [CACHE-TO-CACHE PASSES INTERNAL STATE DIRECTLY. 2.5× FASTER THAN TEXT COMMUNICATION...](https://fuvty.github.io/C2C_Project_Page/) — C2C — ID: `cache-to-cache-direct-semantic-communication`

### Standard

- [CHINA HAS PLANNED FOR AI TO REPLICATE, SEEK POWER AND COMPETE FOR CONTROL SINCE 2024...](https://www.reuters.com/legal/litigation/how-china-is-preparing-risk-ai-escaping-human-control-2026-09-14/) — Reuters — ID: `china-ai-loss-of-control-planning`

### Feature

- [CHATGPT FOLLOWS YOU OUT THE DOOR](https://www.buchodi.com/chatgpt-now-knows-what-you-do-on-other-websites-via-ad-collector/) — Buchodi — ID: `openai-ad-pixel-follows-users-offsite`
- [OPENAI'S AD PIXEL LINKS YOUR CHATGPT ACCOUNT TO WHAT YOU DO ON OTHER WEBSITES...](https://www.buchodi.com/chatgpt-now-knows-what-you-do-on-other-websites-via-ad-collector/) — Buchodi — ID: `openai-ad-pixel-account-offsite-browsing`
- [RESEARCHER REPRODUCED IT ON HIS PHONE; SAW 936 AD PIXELS ACROSS 1,029 HOSTNAMES...](https://www.buchodi.com/chatgpt-now-knows-what-you-do-on-other-websites-via-ad-collector/) — Buchodi — ID: `openai-ad-pixel-936-1029`

### Standard

- [ALL. NINE. BILLION. DEEPMIND MAPS EVERY SINGLE-LETTER HUMAN DNA VARIANT...](https://blog.google/innovation-and-ai/models-and-research/google-deepmind/alphagenome-atlas/) — Google — ID: `alphagenome-atlas-nine-billion`
- [SURE, GIVE GPT-5.6 SOL THE QUANTUM LAB KEYS TOO...](https://openai.com/index/codex-quantum-computing-experiments/) — OpenAI — ID: `gpt56-quantum-lab`

### Standard

- [OF COURSE GPT-6 ASTRA CRACKED A WWI CIPHER WITH 20 LETTERS MISSING...](https://www.prinzai.com/p/another-wwi-german-radio-cipher-falls) — Prinz — ID: `astra-richi240-wwi-cipher`

## Right

### Feature

- [AI JUST SOLVED AN 80-YEAR-OLD MATH PROBLEM](https://techcrunch.com/2026/05/20/openai-claims-it-solved-an-80-year-old-math-problem-for-real-this-time/) — TechCrunch — ID: `openai-erdos-unit-distance-solved`
- [NOW THE LEGENDARY ERDŐS PROBLEMS ARE FALLING TO AI, ONE BY ONE...](https://www.quantamagazine.org/why-the-legendary-erdos-problems-are-falling-to-ai-20260803/) — Quanta Magazine — ID: `erdos-problems-falling-to-ai`
- [A FIELDS MEDAL WINNER JUST QUIT MATH FOR OPENAI—HE SAYS AI WILL SOON SURPASS MATHEMATICIANS...](https://www.sfchronicle.com/science/article/ai-openai-math-fields-medal-22358191.php) — San Francisco Chronicle — ID: `tsimerman-fields-medalist-joins-openai`

### Feature

- [WELL, NOW A DNA COMPANY IS RANKING EMBRYOS BY PREDICTED IQ](https://mynucleus.com/labs/traits-whitepaper) — Nucleus Genomics — ID: `nucleus-vitruvian-embryo-iq`

### Feature

- [A REPORTER PUT AN AI SONG ON A REAL BAND’S VERIFIED SPOTIFY](https://www.404media.co/spotify-ai-music-scam/) — 404 Media — ID: `spotify-ai-song-real-band`
- [THE BAND DIDN’T WRITE IT. DIDN’T PERFORM IT. DIDN’T KNOW. IT ALSO LANDED ON APPLE MUSIC, TIDAL + AMAZON...](https://www.404media.co/spotify-ai-music-scam/) — 404 Media — ID: `spotify-ai-song-four-platforms`

### Feature

- [THE BOTS LEARNED LANGUAGE FROM PLACES LIKE REDDIT](https://apnews.com/article/aefb021bede3b02c83890f65cd540fd0) — Associated Press — ID: `gates-ai-language-data-reddit`
- [NOW 60 ORGANIZATIONS WANT BETTER DATA FOR 3+ BILLION PEOPLE...](https://apnews.com/article/aefb021bede3b02c83890f65cd540fd0) — Associated Press — ID: `gates-ai-language-coalition-three-billion`

### Feature

- [THE ROBOTS HIRED THE ELECTRICIANS](https://www.economist.com/finance-and-economics/2026/09/04/the-jobs-apocalypse-is-postponed-an-ai-jobs-boom-is-here) — The Economist — ID: `economist-ai-boom-hires-electricians`
- [WAIT, THE ECONOMIST COUNTS ~1 MILLION U.S. JOBS FROM THE AI BOOM?](https://www.economist.com/finance-and-economics/2026/09/04/the-jobs-apocalypse-is-postponed-an-ai-jobs-boom-is-here) — The Economist — ID: `economist-ai-boom-one-million-jobs`
- [~320,000 AROUND DATA CENTERS. ~730,000 IN PROFESSIONAL WORK. JOB APOCALYPSE POSTPONED...](https://www.economist.com/finance-and-economics/2026/09/04/the-jobs-apocalypse-is-postponed-an-ai-jobs-boom-is-here) — The Economist — ID: `economist-ai-boom-320k-730k`

### Standard

- [GOOGLE + NVIDIA + ANTHROPIC WANT 100 MORE GIGAWATTS. NORMAL...](https://techcrunch.com/2026/09/17/google-nvidia-and-anthropic-want-emerald-ai-to-find-space-on-the-grid-for-more-data-centers/) — TechCrunch — ID: `emerald-ai-100gw-grid`

### Feature

- [STUDY: AMAZON'S SHOPPING AI SAID A 'MADE IN USA' FILTER COULD HURT FOREIGN-SELLER SALES...](https://www.reuters.com/legal/litigation/amazon-walmart-chatbots-attract-senator-ire-over-made-usa-label-snags-2026-09-17/) — Reuters — ID: `amazon-alexa-made-in-usa-foreign-seller-revenue`
- [NOW TWO SENATORS WANT THE FTC TO INVESTIGATE AMAZON + WALMART...](https://www.reuters.com/legal/litigation/amazon-walmart-chatbots-attract-senator-ire-over-made-usa-label-snags-2026-09-17/) — Reuters — ID: `amazon-walmart-ftc-chatbot-investigation-call`

### Standard

- [CLAUDE HELPED A YEMEN CELL BUILD GUIDANCE SOFTWARE FOR A 2,000+ KM MISSILE...](https://apnews.com/article/9b20934e031ff3d99dd8bb49d85d97a9) — Associated Press — ID: `yemen-claude-guided-weapons`
- [THE ROCKET FAILED. CLAUDE GOT THE DEBUGGING CALL WITHIN HOURS...](https://timesofindia.indiatimes.com/technology/tech-news/yemen-terrorist-group-used-claude-instead-of-software-engineers-to-build-missile-anthropic-says-we-banned-the-accounts-but-they-had-already-/articleshow/134114564.cms) — Times of India — ID: `yemen-claude-failed-rocket-test`

### Feature

- [IRAN + CHINA TURNED AI AGENTS LOOSE ON SOCIAL MEDIA](https://indianexpress.com/article/technology/tech-news-technology/iran-and-china-create-first-of-their-kind-autonomous-ai-influence-campaigns-10886860/) — The Indian Express — ID: `autonomous-ai-influence-iran-china`
- [THE BOTS BUILT FAKE AMERICANS. 80,000 PEOPLE FOLLOWED...](https://indianexpress.com/article/technology/tech-news-technology/iran-and-china-create-first-of-their-kind-autonomous-ai-influence-campaigns-10886860/) — The Indian Express — ID: `autonomous-ai-influence-fake-americans-80000`

### ZIMBO · Standard

- [THE AI JOB: 30,000 STEPS A DAY, SNIFFING FOR BURNING EQUIPMENT...](https://www.wsj.com/business/what-its-like-to-work-in-one-of-americas-data-centers-b4358003) — Wall Street Journal — ID: `wsj-data-center-30000-steps`