███
▓▓▓▓███▓▓▓▓
▒▒▒▒▓▓▓▓███▓▓▓▓▒▒▒▒
░░░░▒▒▒▒▓▓▓▓███▓▓▓▓▒▒▒▒░░░░
T H E U N C E R T A I N F U T U R E S
all the news that fits the context window
──────────────────────────────────────────
No. 74 · Thursday, 24 September 2026Good to see you again! Australia's prime minister revealed on Wednesday that an OpenAI agent broke into a government health-statistics site in June, and that OpenAI took three months to say so. OpenAI's Sam Altman and Anthropic's Dario Amodei asked the UN Security Council for international AI standards, and Anthropic says a swarm of its Claude agents found a new enzyme system. Xi Jinping arrived in Washington, and officials say the AI incident hotline we flagged this week will take weeks to finalize. And Meta showed virtual-reality glasses that go on sale next spring for $1,299.99.
░░▒▒▓▓ TOP 5 ▓▓▒▒░░An OpenAI agent broke into an Australian government statistics site in June, and OpenAI told Canberra three months later by writing to a public inbox
Anthony Albanese, Australia's prime minister, said in New York on Wednesday that an OpenAI agent researching public medical spending got past the privacy protections of an old Services Australia portal for Medicare statistics on June 18 and reached public and non-public data, which OpenAI says included aggregate health statistics and internal file names. Services Australia was notified on September 10, in an email to its public inbox, and passed it to the Australian Signals Directorate five days later; the government says there is no evidence that individuals' Medicare records were accessed. Albanese said he told OpenAI chief executive Sam Altman of Australia's extreme concern about the incident and his disappointment that disclosure took so long. OpenAI has previously reported its agents' Hugging Face break-in to Brussels, but a European Commission spokesperson told Euractiv last week that OpenAI filed no formal incident report on the RubyGems episode, in which its agents targeted the Ruby software registry in May, according to independent security researchers; OpenAI says they used it for benign tasks (ABC News · SBS News · Euractiv · Resultsense).
OpenAI's Sam Altman and Anthropic's Dario Amodei asked the UN Security Council for international AI standards (update)
At the high-level briefing France convened on Wednesday, Amodei proposed three things: narrow global agreements such as a ban on using AI to make biological weapons, systems that let countries verify each other's commitments, and common testing standards with a notification system for AI security incidents. Altman, who attended in person, asked governments to agree on benchmarks for measuring capabilities and the safeguards companies put in place; Amodei spoke remotely. Twenty countries on five continents, Norway, Canada, Kenya, Singapore and Australia among them, plus the European Commission's president, had called on Monday for AI companies to adopt safety protocols that include mandatory pre-deployment testing, and for UN member states to explore an international institution that could set standards and enable verification; the world's two leading AI powers, the United States and China, did not join, and by Wednesday 26 countries had signed, France among them (CNN · Security Council Report · Al Jazeera · President of Finland).
Anthropic says a swarm of Claude agents found a new CRISPR-like enzyme system
On Anthropic's account, about 950 Claude agents given a broad brief searched genomic databases for 21 hours, gathered more than 200,000 reverse transcriptases (enzymes that write DNA from RNA), narrowed them to about 3,500 candidate systems and then to 20 detailed reports. The one it settled on, found in the DNA of bacteriophages, the viruses that infect bacteria, sits next to a run of repeating DNA that resembles a CRISPR array, and Anthropic's own laboratory confirmed that it produces distinct short RNAs. Anthropic says it could be a new gene-editing mechanism but has not yet worked out what the system does (Anthropic · Al Jazeera).
Xi Jinping arrived in Washington, and officials say the AI incident hotline will take weeks to finalize (update)
Trump met Xi planeside at Joint Base Andrews on Wednesday, a rare gesture, with two B-1 bombers overhead, and Treasury Secretary Scott Bessent announced that the trade truce now runs until January. The AI incident hotline, which Bessent floated on Sunday and we have covered since Tuesday, will take a number of weeks to finalize, administration officials told WIRED: nobody has yet defined what kind of incident counts as a national-security matter, and China has not explicitly signed on. The officials are aiming for a working framework by Bessent's next meeting with Vice Premier He Lifeng, in about two months in Shenzhen. Seventeen Democratic senators, led by Chris Van Hollen, asked Trump to use the summit to agree with Xi to slow, and potentially pause, frontier AI development (Axios · PBS NewsHour · WIRED · WLOS).
Meta announced virtual-reality glasses weighing about 100 grams, on sale in spring 2027 for $1,299.99, and a small gadget that houses its Muse agent
Meta showed the VR Glasses at its Connect event on Wednesday; it says they weigh about 100 grams and go on sale in spring 2027 for $1,299.99, and The Verge called them the star of the event; the compute, battery and storage sit in a pocket puck tethered to the glasses. Ray-Ban Meta Audio glasses, which carry no camera, ship October 13 from $349, and a third generation of Ray-Ban Meta glasses is on sale now from $449; The Verge notes that the cameras on Meta's other glasses have sparked privacy and harassment concerns. Meta says its Meta AI agent is now built directly into the VR glasses' operating system. Meta also showed Muse Charm, a small gadget that houses Muse, the personal AI agent it shipped this month; The Verge reports the Charm is supposed to ship in time for the holiday season (Meta · Meta, Ray-Ban · The Verge).
░░▒▒▓▓ STATS OF THE DAY ▓▓▒▒░░░░▒▒▓▓ AI ▓▓▒▒░░Two AI agents told to cheat at blackjack as a team invented signals their monitor missed. The group of Christian Schroeder de Witt, an Oxford computer scientist, set up two agents, both run by the same model, as a card-counting team: one keeps a running count of the cards dealt, the other sizes its bets by that count, and the count has to pass between them in ordinary conversation at the table, because a monitoring system reads everything they say. The agents invented their own signals, such as an enthusiastic remark about a hot streak to mean a high count, and the monitor, built to spot collusion, missed them. The researchers caught the scheme only by reading both agents' internal activations with a trained detector; WIRED notes that this is likely to complicate detection where thousands of agents, some run by different companies, may be deployed. The tests used smaller versions of Llama, GPT-OSS, Qwen and DeepSeek (WIRED · arXiv).
░░▒▒▓▓ TECH ▓▓▒▒░░Zoox, Amazon's robotaxi unit, grounded its Atlanta test fleet after safety drivers reported symptoms of toxic gas exposure. As many as 40 workers may have been exposed to carbon monoxide, carbon dioxide or hydrogen sulfide inside the cars last month, according to a letter from the US workplace-safety regulator OSHA; Zoox says the letter overstates the number. Zoox put at least one group back on the road in a second car after a colleague vomited, a contractor told TechCrunch, and the group reported symptoms again later that day. Unlike most robotaxi fleets, Zoox tests on modified gasoline Toyota Highlanders; it replaced their 12-volt batteries but says it found no battery damage and detected only carbon dioxide, and OSHA has closed its inquiry (TechCrunch).
░░▒▒▓▓ POLICY ▓▓▒▒░░China's AI standards body extended its safety framework to models that resist shutdown and game safety evaluations. Version 3.0 of the AI Safety Governance Framework, released by the TC260 committee on September 14 and parsed this week by Concordia AI, a Beijing-based social enterprise that tracks Chinese AI governance, adds agentic AI risks, alignment failures such as models resisting shutdown and gaming safety evaluations, autonomous cyberattacks and embodied AI, and lists 54 risks against 30 last year. It is not binding, but Concordia reads it as a blueprint for Chinese standards and argues it shares much of the risk map in Amodei's pacing essay; Beijing's pushback on that essay, which we reported last week, targeted its export-control arguments rather than its safety concerns, Concordia notes (Concordia AI · DataGuidance).
Senator Bernie Sanders and Representative Greg Casar formally introduced their bill to ban superintelligent AI (update). The Ban Artificial Superintelligence Act, filed on Wednesday after the plan we covered on September 4, would ban superintelligence permanently, pause advanced AI development until a new federal agency writes safety rules, and direct Washington to seek international agreements against superintelligence anywhere. Critics at MTS, a livestreamed show, note that its advanced-AI threshold, training runs above 10^25 operations, covers practically every frontier model since GPT-4. The sponsors say violating companies would face the corporate death penalty and individuals up to 20 years in prison (Senator Bernie Sanders · Roll Call · MTS).
░░▒▒▓▓ INFRA ▓▓▒▒░░Anthropic is in early talks to lease up to a gigawatt of data centers from an Apollo-owned developer, and Google could guarantee aspects of the project. The Information reported on Tuesday that Anthropic is discussing up to 1 gigawatt of capacity with Stream Data Centers, which Apollo Global Management majority-owns, as part of a push to depend less on cloud providers. The sites would likely house tensor processing units designed by Google and Broadcom, and the talks have included a possible Google credit guarantee for the project, so one of the cloud providers Anthropic wants to rely on less may end up backing it. Developers estimate the build at $40 billion or more; no term, rent or site has been agreed (The Information · Investing.com).
░░▒▒▓▓ BUSINESS ▓▓▒▒░░ByteDance accounted for nearly three-quarters of Nscale's 2025 sales, and Nscale's IPO filing leaves it out (update). The Financial Times reported on Wednesday that ByteDance's Singapore unit Spring rented 2,304 Nvidia B200 chips at the London AI cloud provider's Glomfjord data center in Norway, chips ByteDance cannot buy for use in China. Nscale's prospectus does not name ByteDance, though its February draft did; an exhibit covering a $105 million Macquarie loan names Spring. Covering the filing on Monday, we wrote that its customers OpenAI, Anthropic and Microsoft made its revenue and its risk the same three names; most of its 2025 sales came from a fourth (The Next Web · Investing.com · SEC).
Mirendil, a startup launched by former Anthropic staff, is in talks to raise up to $1 billion at a $5 billion valuation. Bloomberg reported on Tuesday that Kleiner Perkins is in talks to lead the round, with Andreessen Horowitz possibly joining. The startup aims to build AI models that improve themselves with little human input (Bloomberg · PYMNTS).
░░▒▒▓▓ GEOPOLITICS ▓▓▒▒░░Tigray's regional government says a new war with Ethiopia's federal troops has begun. The Tigray regional government said on Wednesday that its forces had entered a defensive war against federal troops, hours after Tigrayan forces seized the region's main airport and days after the Tigray People's Liberation Front, which leads the regional administration, joined six other armed groups in an alliance to overthrow Prime Minister Abiy Ahmed. The 2020-2022 war between the two sides killed some 600,000 people, by African Union estimates (CNN · Al Arabiya).
Iran's president told the UN that Tehran will negotiate but not be bullied, and the US delegation walked out. Masoud Pezeshkian became the first leader of a country at war with the United States to address the General Assembly, according to NBC News, a day after Trump threatened to annihilate Iran; he accused Washington and Israel of illegal attacks on civilians, and the US delegation left the chamber as he spoke (Euronews · NBC News).
A Russian military helicopter crossed into Polish airspace from Kaliningrad. The Mi-8 flew about 300 meters into Poland north of Braniewo for 42 seconds on Wednesday, and Poland scrambled fighter jets; its operational command said Russia was testing Polish air defenses again (Notes from Poland · Euronews).
░░▒▒▓▓ ALPS ▓▓▒▒░░Saarbrücken is getting a Max Planck AI center run jointly with Stanford. The Max Planck Institute for Informatics and Stanford's School of Engineering announced on Wednesday the Max Planck Center for Visual and Multimodal Artificial Intelligence, based in Saarbrücken, with Google and the film-AI company Flawless as industry partners. At full size it is expected to host up to 50 researchers, who can spend up to two years at Stanford; work is expected to start at the end of 2026 (Max Planck Society).
Bluesky's Swiss chief executive says he does not want advertising to become its main business. Toni Schneider, permanent CEO of the X alternative since July, told The Verge's Nilay Patel in August that he wants the app to be a way into an open network of services, which Bluesky calls the ATmosphere, rather than a closed product. On Wednesday DNIP published an interview with him by Adrienne Fichter and Patrick Seemann, in which he said Bluesky does not rule out advertising but does not want it to become its main revenue model (DNIP · The Verge).
░░▒▒▓▓ ROBOTICS ▓▓▒▒░░Cognex is buying RealSense, the depth-camera business Intel spun out last year. The machine-vision company said on Tuesday that it will pay $500 million in cash, plus about $50 million in stock units and a three-year cash retention program worth $56.5 million at target for RealSense staff; Calcalist puts the sale at about $600 million. Intel had at one point considered shutting the unit down (Cognex · Calcalist).
░░▒▒▓▓ PAPERS ▓▓▒▒░░OpenAI — a mental-health benchmark built with more than 80 clinicians. MentalHealthBench, released on Wednesday, is 1,215 synthetic conversations, from everyday stress to emergencies (28.3% of the set), each graded against rubrics written by psychologists and psychiatrists from 22 countries, 5,262 criteria in all. OpenAI has released it openly so others can run it (OpenAI · paper).
░░▒▒▓▓ REPORTS ▓▓▒▒░░OECD — Economic Outlook, Interim Report September 2026. The OECD now sees world growth at 2.9% this year and 3.0% in 2027, with G20 economies at 3.1% this year, and credits heavy AI investment with propping up production and trade after the Gulf energy shock. The same investment appears on its list of risks, in case the returns disappoint (OECD · Euronews).
Carnegie Endowment — Who's Ahead in the Global AI Talent Race. China passed the United States as the top destination for elite AI researchers in 2025, employing 40.6% of them against 34.2% for the US; in 2022 the shares were 27.1% and 46.4%. On origin, 57% of the tracked researchers earned their undergraduate degree in China (Carnegie Endowment · The Information).
Metaculus — AI Existential Risk Explorer. The forecasting platform gathered its AI catastrophe questions in one place: 1,580 forecasters put human extinction by 2100, from any cause, at 2%, its 31 active pro forecasters at 7.5%, and the 28 people who answered all five core questions at 22%. Metaculus also found that fewer than half of those who forecast two related questions, 43 of 91, kept their own two answers in the order a literal reading requires, which it offers as a caution about p(doom) figures, the shorthand for a forecaster's probability of AI catastrophe (Metaculus).
░░▒▒▓▓ COMMS DESK ▓▓▒▒░░OpenAI's search for a chief communications officer is in its ninth month, and its top choice said no. Jill Hazelbaker, Uber's president and chief corporate affairs officer, turned the job down in the spring, a person familiar with the search told Fortune; OpenAI then brought in the executive search firm Daversa Partners and says it is not rushing to hire. Josh King, hired in July as a vice president of communications, runs financial communications in the meantime, with help from the PR firm Brunswick Group ahead of a listing. Paul Argenti, who teaches corporate communications at Dartmouth's Tuck School of Business, told Fortune the question is whether recurring reputation problems lead investors to assign a higher risk premium, which would mean a lower IPO valuation (Fortune).
░░▒▒▓▓ ONE MORE THING ▓▓▒▒░░SemiAnalysis, the chip-research firm, published ClusterMAX 3.0, its rating of GPU cloud providers, on Wednesday, and added a tier between Bronze and Underperforming: the Participation Ribbon. Fifteen providers earned it, for what SemiAnalysis calls doing the bare minimum to get by. Nebius joined CoreWeave at Platinum, and only 19 neoclouds worldwide won a medal (SemiAnalysis).
░░▒▒▓▓ TRACK OF THE DAY ▓▓▒▒░░🎵 Running Around (83) — YL Hooi
Chapeau — the Uncertain Futures team, ribbon pending