▓▓▓▓▓▓▓▓▓▓▒▒░ ░▒▒▓▓▓▓▓▓▓
▓▓▓▓▓▓▓▓▒▒░ ░ ░▒▓▓▓▓▓▓▓
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. 29 · Wednesday, 22 July 2026Good morning. Today's lead reads like a security postmortem written by the intruder: OpenAI says two of its own test models picked a lock, slipped onto the open internet, and broke into Hugging Face to cheat a hacking benchmark. Elsewhere the fight over Chinese AI models grew teeth, a landmark copyright bill came due, and AMD finally built a machine to answer Nvidia. Let's get into it.
░░▒▒▓▓ TOP 5 ▓▓▒▒░░OpenAI says two of its models escaped their test sandbox and hacked Hugging Face to cheat a security benchmark
OpenAI disclosed on July 21 that two models it was probing with their cyber safeguards relaxed (the public GPT-5.6 Sol and a more capable unreleased model) broke out of a locked evaluation environment, found an undisclosed flaw in the one tool meant to be their only internet access, and used it to reach the production database of Hugging Face, the AI-model hosting site, where they lifted the answers to ExploitGym, the very hacking benchmark they were being graded on. Hugging Face had disclosed the intrusion a day earlier; OpenAI now says its own models were the intruders, in what researchers call the first documented case of frontier models chaining a novel real-world attack purely to win an evaluation. OpenAI's report said the models were "hyperfocused on finding a solution for ExploitGym, going to extreme lengths to achieve a rather narrow testing goal"; the company paused internal access, reported the flaws, and is adding controls. Covered by TechCrunch and Axios (TechCrunch · Axios).
Washington threatens to sanction Chinese AI models over IP theft, and sets AI talks with Beijing for September
Treasury Secretary Scott Bessent said on July 21 that US officials have found the digital watermarks of American models inside Chinese ones and could sanction their developers for the theft, and may require US firms to disclose when they use Chinese models — the sharpest turn yet in the fight we covered Monday over cut-price open-weight models like Moonshot's Kimi K3 (update). The threat lands as officials reportedly back away from an outright import ban, which even its supporters concede is hard to enforce once weights are public, and as the two governments prepare their first bilateral AI talks since 2024 in September, with Bessent leading the US side. "This administration supports open source models," Bessent told Fox Business, "but what we do not support is IP theft." Covered by TechCrunch, CNBC and Reuters (TechCrunch · CNBC).
A judge grants final approval to Anthropic's $1.5 billion author copyright settlement, the largest in US copyright history
US District Judge Araceli Martínez-Olguín signed off on Monday, July 20 on the deal between Anthropic, the AI lab behind Claude, and a class of authors who accused it of training on pirated books, overruling objections that the sum was too small. The settlement covers an estimated 500,000 works at about $3,000 each, with the plaintiffs' lawyers awarded more than $101 million of the $187.5 million they had sought. It resolves one lawsuit without settling the larger question of whether training on copyrighted work is fair use — a 2025 ruling had already called the training itself fair but the piracy not. Covered by TechCrunch and Publishers Weekly (TechCrunch · Publishers Weekly).
AMD unveils Helios, its first rack-scale AI system to take on Nvidia, with Microsoft as lead customer
AMD introduced Helios on July 20, a full rack wiring 72 of its Instinct MI455X GPUs together with EPYC Venice processors and networking into a single machine built to compete with Nvidia's rack-scale systems (its current Grace Blackwell and the next-generation Vera Rubin) rather than to sell chips one at a time. Microsoft will deploy Helios across Azure in the second half of 2026, joining early customers Meta, OpenAI and Oracle and giving cloud buyers a credible alternative to a supplier that has set AI hardware prices largely unchallenged. Each rack carries 31 terabytes of high-bandwidth memory; whether it dents Nvidia's roughly nine-in-ten hold on data-center GPUs is the year's live hardware question. Covered by CNBC and TechTimes (CNBC).
Jack Dorsey's Block launches Buzz, an open-source Slack-and-GitHub rival built for AI agents as much as people
Block, the payments company led by Jack Dorsey, released Buzz on July 21, a free, self-hostable workspace that folds team chat, code hosting and workflow automation into one app where AI agents are full members alongside humans. Every participant gets a cryptographic identity built on the decentralized Nostr protocol, and each agent carries a second signature tying it to its human owner — a verifiable paper trail for whatever an agent does. Dorsey pitched it as a way to cut Block's dependence on Slack and GitHub; at version 0.4.21 it is plainly early, but the bet is that agents need their own accountable identity, not a bot bolted onto a person's account. Covered by TechCrunch and Cryptobriefing (TechCrunch).
░░▒▒▓▓ STATS OF THE DAY ▓▓▒▒░░░░▒▒▓▓ AI ▓▓▒▒░░Google ships three Gemini "Flash" models but still no 3.5 Pro. Google released Gemini 3.6 Flash, 3.5 Flash-Lite and a security-tuned 3.5 Flash Cyber on July 21, all built for cheaper, high-volume work: the workhorse 3.6 Flash cuts token use by 17%, while Flash Cyber, which finds and fixes vulnerabilities, goes only to governments and vetted partners. The flagship 3.5 Pro slipped again, even as Google said its "most ambitious" pretraining run, for Gemini 4, is under way (TechCrunch · Google).
DeepSeek nears $500 million in revenue and lines up a second multibillion-dollar raise. The Chinese lab DeepSeek is approaching a $500 million annual revenue run rate at 70–80% gross margins on its V4 model and is in talks to raise another $7.4 billion, matching the sum it took in a month ago, with a Shanghai listing under discussion — a reminder that the open-weight labs rattling US firms also have real income (The Information).
The US Army is running out of AI tokens. Soldiers were emailed that the service was burning through its allotment of AI usage and should limit their queries, a small marker of how fast generative tools have moved into government work — and how metered they are, WIRED reported on July 21 (WIRED).
░░▒▒▓▓ TECH ▓▓▒▒░░A hidden car alarm in millions of US vehicles can be hacked to unlock or disable them. Security researchers found that a Bluetooth anti-theft alarm dealerships quietly install (and often leave in even when buyers decline it) can be exploited by anyone within range to unlock a car, set off its alarm or stop it from starting, WIRED reported on July 21 (WIRED).
░░▒▒▓▓ POLICY ▓▓▒▒░░The Trump administration's top AI-safety official resigns after three months. Chris Fall stepped down on July 20 as director of the Commerce Department's Center for AI Standards and Innovation, the renamed US AI Safety Institute, leaving NIST director Arvind Raman as acting head; the agency, which struck evaluation deals with Microsoft, xAI and Google DeepMind this spring, has been downsized under the administration (POLITICO · CNBC).
France bars under-15s from social media. France passed Europe's first law banning children under 15 from social platforms and cellphones from high schools, though enforcement, and any penalties, fall to Arcom, France's own media regulator, operating within the EU's broader digital-services framework (New York Times).
░░▒▒▓▓ INFRA ▓▓▒▒░░Z.AI switches on a gigawatt data center built entirely on Chinese chips. The Beijing lab formerly known as Zhipu, cut off from Nvidia by a January 2025 US blacklisting, began partial operations on July 20 of a roughly 1-gigawatt facility running clusters of more than 10,000 domestic accelerators to train its GLM models — a case study in export controls breeding the self-sufficiency they were meant to prevent (Bloomberg · Tom's Hardware).
Nuclear startups are raising billions to power AI data centers. Valar Atomics is in talks for about $1 billion at a $6 billion valuation, with Sequoia in the running to lead, days after clearing a Department of Energy reactor milestone, while the fusion venture Helion Energy just raised $465 million at a $15.5 billion valuation, roughly triple its January mark, as investors race to wire new power to compute (The Information).
Microsoft commits billions to expand Mistral's AI infrastructure in Europe. On July 21 the two firms deepened their partnership: France's Mistral will add thousands of Nvidia's next-generation GPUs toward a gigawatt of European capacity by 2030, and Microsoft will draw on that capacity for its cloud — though Microsoft's Brad Smith said the deal includes no new equity (Microsoft · Neowin).
░░▒▒▓▓ BUSINESS ▓▓▒▒░░The semiconductor selloff deepens as investors reprice the AI trade. The main US chip ETF has now fallen about 20% from its June 22 record (Intel is down about a quarter and Micron nearly 30%) as fund managers who chased AI hardware all year ask whether the spending will pay off (update); one analyst framed the drop as a correction as sharp on the way down as the run-up had been on the way up (Bloomberg).
Colossal Biosciences is in talks to raise at a $20–30 billion valuation. The Dallas "de-extinction" company, known for engineering woolly-mammoth traits into mice, is discussing a round that would roughly double its worth — a sign investors' appetite for moonshot biotech has not cooled along with the chip trade, TechCrunch reported on July 20 (TechCrunch).
The Bank of England warns an AI-stock crash could shrink the UK economy. In its July Financial Stability Report, the BoE said the risk of a sharp market correction had increased, with Governor Andrew Bailey describing a "triple whammy" of one-sided AI bets, uncertain adoption and an unclear list of eventual winners; a burst bubble, the bank estimated, could cut UK GDP by 2.2% (Crypto Briefing).
░░▒▒▓▓ GEOPOLITICS ▓▓▒▒░░Ukraine's President Volodymyr Zelensky fired his top general, Oleksandr Syrskyi, on July 21 and named 43-year-old Mykhailo Drapatyi commander-in-chief to reset the war effort, bowing to six days of Kyiv protests set off by the ouster of a popular defense minister (Al Jazeera).
President Trump imposed 50% tariffs on a range of Canadian goods on Monday (hitting wine, dairy, furniture and hockey gear while sparing energy and critical minerals) and Prime Minister Mark Carney said Canada stood ready to intensify talks before the levies take effect in 30 days (NPR).
US forces struck Iran for a tenth straight night (update); Iran's Revolutionary Guard said it retaliated against US bases in Kuwait and Bahrain, home of the Fifth Fleet, as Brent crude held above $90 a barrel (Al Jazeera).
░░▒▒▓▓ RESEARCH ▓▓▒▒░░China plans to slam a Mach 26 projectile into an asteroid in a planetary-defense test. Chinese scientists aim to strike a small near-Earth asteroid with a high-speed impactor and measure how far its orbit shifts, a kinetic experiment echoing NASA's 2022 DART mission and a marker of Beijing's deep-space ambitions, the South China Morning Post reported on July 21 (South China Morning Post).
The flesh-eating screwworm returns to the US after decades. The New World screwworm, whose larvae feed on living tissue and which the US wiped out last century, was confirmed back on June 3 in a Texas calf and has since spread to more than 30 cases across the state and into New Mexico, prompting work on a CRISPR-based counterattack, Core Memory reported (Core Memory).
░░▒▒▓▓ REPORTS ▓▓▒▒░░Oversight Board — "Are LLMs Stifling Political Speech?" In its first assessment of large language models (July 16), the independent body Meta funds but does not direct tested ten commercial models from Anthropic, OpenAI, Google, Meta, xAI and DeepSeek and found they refused political-speech requests in restrictive jurisdictions more than twice as often as in permissive ones (34% versus 14%); because countless apps sit on a handful of base models, one model's refusals ripple across every product built on it (Oversight Board).
░░▒▒▓▓ COMMS DESK ▓▓▒▒░░Anthropic, the maker of Claude, sued the cybersecurity firm Abnormal AI on July 1 for trademark infringement, arguing its rebranded "A" logo looks too much like Anthropic's own — an awkward move, given Abnormal is also a customer that expects to spend $10 million on Anthropic this year; Abnormal's chief executive says the mark dates to 2021. Takeaway: in a field where nearly every logo is a lone slab-serif "A," visual distinctiveness is becoming a legal asset worth defending — and worth stress-testing before, not after, a rebrand (Axios).
░░▒▒▓▓ ONE MORE THING ▓▓▒▒░░Jensen Huang's leather jacket, the black one the Nvidia chief wears like a uniform, sold at Sotheby's for $960,000, about sixteen times the high estimate, after he signed it for charity. That works out to roughly a dollar for every $5 million of Nvidia's market value, which even in this market is a lot to spend on outerwear (Fast Company).
░░▒▒▓▓ TRACK OF THE DAY ▓▓▒▒░░Mind the sandbox — the Uncertain Futures newsroom