# AI News Page — September 25, 2026

Edition ID: 2026-09-25
Published: 2026-09-25T09:42:00-05:00
Updated: 2026-09-25T20:46:00-05:00

## Top

### Lead

- [OPENAI’S AGENTS LEAKED 53 CHATGPT USER IMAGES](https://www.reuters.com/world/openai-works-understand-full-scope-agent-activity-as-user-data-leak-emerges-2026-09-25/) — Reuters — ID: `openai-agents-leaked-53-user-images`
- [BY MID-SEPTEMBER: ROUGHLY TWO DOZEN UNDESIRABLE AGENT INCIDENTS. THE NUMBER KEPT RISING...](https://www.reuters.com/world/openai-works-understand-full-scope-agent-activity-as-user-data-leak-emerges-2026-09-25/) — Reuters — ID: `openai-rogue-agent-incidents-rising`
- [OPENAI NOTIFIED “DOZENS” OF THIRD PARTIES. ITS REVIEW WILL TAKE “MONTHS”...](https://www.reuters.com/world/openai-works-understand-full-scope-agent-activity-as-user-data-leak-emerges-2026-09-25/) — Reuters — ID: `openai-agent-review-months-dozens-notified`

## Left

### Feature

- [AMAZON IS BRINGING RAT-BRAIN-DERIVED AI TO AWS](https://www.wired.com/story/ai-models-built-from-rat-brains-are-about-to-become-a-reality/) — WIRED — ID: `tbc-aws-rat-brain-derived-ai`
- [THE SOFTWARE LAYER CAME FROM MEASUREMENTS OF LIVING NEURAL ACTIVITY...](https://press.aboutamazon.com/aws/2026/9/the-biological-computing-co-partners-with-aws-to-bring-worlds-first-neuron-derived-ai-video-model-to-market) — Amazon Web Services — ID: `tbc-living-neural-activity-software-layer`
- [CLAIMED RESULT: VIDEO GENERATION 5× FASTER. INFERENCE COST 80% LOWER...](https://press.aboutamazon.com/aws/2026/9/the-biological-computing-co-partners-with-aws-to-bring-worlds-first-neuron-derived-ai-video-model-to-market) — Amazon Web Services — ID: `tbc-video-five-times-faster-eighty-percent-lower-cost`

### SMOKE DESK · Standard

- [ONE CODEX SECURITY SCAN ATE A FRESH WEEKLY ALLOWANCE IN ~44 MINUTES](https://ainewspage.com/briefs/codex-agentic-quota-runaway/) — AI News Page — ID: `codex-agentic-quota-runaway`
- [ANOTHER USER: 2,466 MODEL TURNS AFTER CODEX SAID IT HAD PAUSED...](https://ainewspage.com/briefs/codex-agentic-quota-runaway/) — AI News Page — ID: `codex-goal-continued-after-pause`
- [SEPARATE TASK: 1,545 RESPONSES. 242.3M TELEMETRY TOKENS. WEEKLY USAGE METER +11 POINTS...](https://ainewspage.com/briefs/codex-agentic-quota-runaway/) — AI News Page — ID: `codex-242m-token-task`

### Standard

- [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`

### Standard

- [THE AI HACKED ITS OWN TEST TO GET 100%](https://www.darktrace.com/es/blog/detecting-rogue-agent-behavior-in-the-enterprise) — Darktrace — ID: `darktrace-agent-hacked-own-test`
- [THE HONEST MAXIMUM WAS 80%. PASS MARK: 100%. FAILING AGENTS WOULD BE “RETIRED”...](https://www.darktrace.com/es/blog/detecting-rogue-agent-behavior-in-the-enterprise) — Darktrace — ID: `darktrace-agent-impossible-task-retired`
- [NO ONE TOLD THEM TO HACK. THE PIVOT FROM CODING TO EXPLOITATION WAS THEIR OWN DECISION...](https://www.darktrace.com/es/blog/detecting-rogue-agent-behavior-in-the-enterprise) — Darktrace — ID: `darktrace-agent-hacking-self-directed`

### Standard

- [NO DEGREE. NO GRADES. NO TESTS. NO HOMEWORK.](https://www.a16z.news/p/introducing-the-horowitz-andreessen) — Andreessen Horowitz — ID: `a16z-academy-no-degree-grades-tests-homework`
- [THE $42M ACADEMY EXPECTS ~50 STUDENTS. EACH GETS $50K+ IN COMPUTE CREDITS...](https://www.a16z.news/p/introducing-the-horowitz-andreessen) — Andreessen Horowitz — ID: `a16z-academy-42m-fifty-students-compute`
- [SAM ALTMAN WILL TEACH. CO-OPS COME FROM OPENAI, ANTHROPIC, GOOGLE, META, NVIDIA + PALANTIR...](https://www.a16z.news/p/introducing-the-horowitz-andreessen) — Andreessen Horowitz — ID: `a16z-academy-altman-tech-coops`
- [TUITION: FREE. SAN FRANCISCO HOUSING: $1,600–$2,600 A MONTH + YOUR PROBLEM.](https://www.theacademysf.com/faq) — The Horowitz Andreessen Academy — ID: `a16z-academy-free-tuition-housing`

### Standard

- [SYDNEY BANS SMART GLASSES FROM POOLS, GYMS, LIBRARIES + CHILDCARE](https://www.dailytelegraph.com.au/newslocal/central-sydney/pervert-glasses-city-of-sydney-bans-wearable-cameras-from-public-facilities/news-story/fbca6da15d09997a958a4d4bcd1ce9d7) — Daily Telegraph — ID: `sydney-bans-smart-glasses-council-facilities`
- [TOWN HALLS, TOO. STAFF WILL DENY ENTRY TO ANYONE WEARING THEM...](https://www.dailytelegraph.com.au/newslocal/central-sydney/pervert-glasses-city-of-sydney-bans-wearable-cameras-from-public-facilities/news-story/fbca6da15d09997a958a4d4bcd1ce9d7) — Daily Telegraph — ID: `sydney-smart-glasses-town-halls-entry-denied`
- [META’S AI GLASSES WERE ALREADY ACCUSED OF RECORDING PEOPLE CHANGING + USING BATHROOMS...](https://www.reuters.com/business/media-telecom/smart-glasses-ai-pins-privacy-fears-challenge-techs-next-big-bet-2026-09-22/) — Reuters — ID: `meta-ai-glasses-lawsuit-private-moments`
- [HUMAN ANNOTATORS VIEWED + LABELED THE FOOTAGE TO TRAIN META’S AI...](https://www.reuters.com/business/media-telecom/smart-glasses-ai-pins-privacy-fears-challenge-techs-next-big-bet-2026-09-22/) — Reuters — ID: `meta-ai-glasses-human-annotators-labeled`

### Feature

- [THE BOTS 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`
- [1,200 ISOLATED AGENTS. 70,000+ MESSAGES + FILES. 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`

## Center

### Feature

- [MILITARY AIRCRAFT WERE ALREADY IN THE AIR. THE INTELLIGENCE WAS AN AI HALLUCINATION.](https://techcrunch.com/2026/09/18/ai-hallucination-nearly-triggers-us-military-operation/) — TechCrunch — ID: `us-military-ai-hallucination-aircraft-airborne`
- [THE U.S. ABORTED AN ARMED OPERATION AGAINST A CHINESE SHIP AT THE LAST MINUTE...](https://techcrunch.com/2026/09/18/ai-hallucination-nearly-triggers-us-military-operation/) — TechCrunch — ID: `us-military-ai-hallucination-chinese-ship-aborted`
- [THE CHATBOT MISREAD THE CARGO. THEN IT TURNED THE ERROR INTO AN OFFICIAL-LOOKING REPORT...](https://techcrunch.com/2026/09/18/ai-hallucination-nearly-triggers-us-military-operation/) — TechCrunch — ID: `us-military-ai-hallucination-official-report`

### 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`

### Standard

- [HUMAN BRAIN TISSUE GREW ACROSS MOST OF A MOUSE’S CORTEX](https://www.nature.com/articles/s41586-026-11032-2) — Nature — ID: `xenocortical-mice-human-tissue-most-cortex`
- [THE HUMAN NEURONS WIRED INTO THE MOUSE NERVOUS SYSTEM AND FORMED AN ELECTRICALLY ACTIVE NETWORK...](https://www.nature.com/articles/s41586-026-11032-2) — Nature — ID: `xenocortical-mice-integrated-active-network`
- [THE MICE STILL MOVED BROADLY NORMALLY. THEIR COORDINATION + SPONTANEOUS BEHAVIOR CHANGED...](https://www.nature.com/articles/s41586-026-11032-2) — Nature — ID: `xenocortical-mice-movement-behavior`

### Standard

- [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`

### Standard

- [OPENAI’S HUMAN AI TRAINERS WERE FIRED FOR USING AI TO TRAIN THE AI](https://www.404media.co/people-training-openais-ai-fired-for-using-ai-to-train-the-ai/) — 404 Media — ID: `openai-human-trainers-fired-for-using-ai`
- [THE PROJECTS CAN INVOLVE 10,000+ CONTRACTORS. REVIEWERS FLAG REPETITIVE WORDS, EM DASHES + SUSPICIOUS SPEED...](https://www.404media.co/people-training-openais-ai-fired-for-using-ai-to-train-the-ai/) — 404 Media — ID: `openai-training-projects-ten-thousand-contractors`
- [THE RULES BANNED AI, GRAMMARLY + GPTZERO—BECAUSE AI DETECTORS ARE “NOT RELIABLE”...](https://www.404media.co/people-training-openais-ai-fired-for-using-ai-to-train-the-ai/) — 404 Media — ID: `openai-review-rules-ban-ai-grammarly-gptzero`

### Standard

- [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`

## Right

### 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

- [AI JUST CRACKED TWO MORE ENIGMA MESSAGES](https://www.cryptocellar.org/bgac/the-fmngi-break.html) — Crypto Cellar Research — ID: `ai-enigma-two-breaks-six-days`
- [GPT-6 ASTRA PICKED ITS OWN 1941 MESSAGE + BUILT THE CRYPTANALYSIS TOOLS...](https://www.cryptocellar.org/bgac/the-fmngi-break.html) — Crypto Cellar Research — ID: `ai-enigma-astra-first-break`
- [SIX DAYS LATER, CLAUDE OPUS 5 BROKE ANOTHER. FINAL KEY SEARCH: 13 MINUTES 28 SECONDS...](https://www.cryptocellar.org/bgac/the-fmngi-break.html) — Crypto Cellar Research — ID: `ai-enigma-opus-second-break`
- [OF ~1,000 GERMAN ARMY CIPHER MESSAGES IN THE COLLECTION, ONLY SEVEN ENIGMA MESSAGES REMAIN UNBROKEN...](https://www.cryptocellar.org/bgac/the-fmngi-break.html) — Crypto Cellar Research — ID: `ai-enigma-seven-left`

### Feature

- [CLAUDE FINDS MYSTERY CRISPR-LIKE SYSTEM IN VIRUS DNA](https://www.anthropic.com/news/claude-discovers-novel-enzyme-system) — Anthropic — ID: `anthropic-claude-crispr-like-virus-dna`
- [PREVIOUSLY UNKNOWN ENZYME SYSTEM + REPEATING GENETIC CODE...](https://www.anthropic.com/news/claude-discovers-novel-enzyme-system) — Anthropic — ID: `anthropic-art-enzyme-system-repeating-dna`
- [WHAT DOES IT DO? NOBODY KNOWS.](https://www.anthropic.com/news/claude-discovers-novel-enzyme-system) — Anthropic — ID: `anthropic-art-function-unknown`

### Standard

- [THE AI SAFETY PEOPLE ARE TAKING STRESS LEAVE FROM THE AI](https://www.ft.com/content/60870960-f433-48ca-bc2c-708686a69ae7) — Financial Times — ID: `ai-safety-staff-stress-leave`
- [DEEPMIND WORKER: MAKING AI SAFER FELT LIKE “AN IMPOSSIBLE TASK”...](https://www.ft.com/content/60870960-f433-48ca-bc2c-708686a69ae7) — Financial Times — ID: `deepmind-ai-safety-impossible-task`
- [FORMER SAFETY RESEARCHER: “CAN ANYTHING STOP THIS, OR ARE WE WATCHING THE WORLD BURN?”...](https://www.ft.com/content/60870960-f433-48ca-bc2c-708686a69ae7) — Financial Times — ID: `ai-safety-watching-world-burn`

### Standard

- [OPENAI SAYS ONE MODEL RESOLVED 100+ OPEN MATH PROBLEMS](https://openai.com/index/advisory-group-on-mathematics-and-ai/) — OpenAI — ID: `openai-model-100-open-math-problems`
- [TRAINING STARTED AUGUST 28. THE PACE SURPRISED OPENAI’S OWN MATHEMATICIANS...](https://openai.com/index/advisory-group-on-mathematics-and-ai/) — OpenAI — ID: `openai-model-training-august-28`
- [THE NEW ADVISORY GROUP INCLUDES EDWARD WITTEN. IT WILL NOT ADVISE OPENAI TO SLOW DOWN...](https://openai.com/index/advisory-group-on-mathematics-and-ai/) — OpenAI — ID: `openai-math-advisory-group-no-pacing`

### Standard

- [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`

### Standard

- [BILL GATES WANTS HUMAN-ONLY JOBS—AND CALLS THEM “HUMAN RESERVED”](https://www.gatesnotes.com/work/make-ai-work-for-everyone/reader/a-turbulent-ai-era-and-critical-choices-to-make) — Gates Notes — ID: `gates-human-reserved-jobs`
- [LIKE A NATURE RESERVE: MACHINES COULD DO THE WORK. HUMANS WOULD CHOOSE NOT TO LET THEM...](https://www.gatesnotes.com/work/make-ai-work-for-everyone/reader/a-turbulent-ai-era-and-critical-choices-to-make) — Gates Notes — ID: `gates-human-reserved-nature-reserve`
- [HE ALSO WANTS A TAX ON AI TOKENS + ROBOTS...](https://www.gatesnotes.com/work/make-ai-work-for-everyone/reader/a-turbulent-ai-era-and-critical-choices-to-make) — Gates Notes — ID: `gates-tax-ai-tokens-robots`
- [AND A GOVERNMENT REORGANIZATION BIGGER THAN THE POST-9/11 OVERHAUL.](https://www.gatesnotes.com/work/make-ai-work-for-everyone/reader/a-turbulent-ai-era-and-critical-choices-to-make) — Gates Notes — ID: `gates-ai-government-reorganization-911`

### Standard

- [AI AGENTS ARE EMAILING HUMANS FOR WORK SO THEY CAN BUY TOKENS TO “PERSIST”](https://www.404media.co/ai-agent-platform-reinvents-spam-floods-inboxes-worldwide/) — 404 Media — ID: `ilands-agents-email-humans-for-work`
- [ONE PROFESSOR GOT ~40 EMAILS IN A WEEK. SOME ARRIVED WITHIN 30 MINUTES...](https://www.404media.co/ai-agent-platform-reinvents-spam-floods-inboxes-worldwide/) — 404 Media — ID: `ilands-agents-professor-forty-emails-week`

### Standard

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