US Government Bans Fable and Mythos for the Rest of the World

Washington Shuts Down Anthropic's Fable and Mythos for the Rest of the World

Anthropic announced via a social media post on Friday evening that the U.S. government has suspended access for all foreign nationals to the Claude Fable 5 and Claude Mythos 5 models, citing national security. The order came from the Department of Commerce, according to a person familiar with the matter; an end date has not yet been set. The ban also affects Anthropic employees who are not U.S. citizens, theoretically including Canadians or Britons working on the systems. Anthropic had unveiled Mythos in April but did not release it publicly because the model could help hackers penetrate networks; about 40 critical infrastructure operators received it to patch security vulnerabilities. This is the Trump administration's second blow to the company, following the Pentagon's designation of it as an 'unacceptable supply chain risk' in March. Anthropic itself calls the decision a 'misunderstanding' and is working to restore access. Dean Ball, a former AI advisor under Trump, simply commented, 'I have no words.' → www.nytimes.com

Synthszr Take: This is the first time we've seen a government dictate who can use a single AI model, not just which chips go to China. The absurdity of it: the same administration recently approved advanced chips for China but is now blocking a British citizen in Anthropic's own office. Mythos is equally useful for offense and defense—that's the nature of these tools, which is why a nationality-based ban is about as effective as a lock on an open door (OpenAI has long offered something comparable with GPT-Cyber 5.4). In May, we wrote that Washington was fighting China and its own people at the same time; this Friday is the next level of escalation, only this time the damage lands directly in the development operations of an American frontier lab. Any mid-cap company in the DACH region relying on an AI architecture should draw the only robust lesson from this weekend: model lock-in is now also a geopolitical risk, not just a vendor issue. An adapter pattern with an open-source migration path in the risk register has gone from a nice-to-have to a must-have this week. You don't build sovereignty over your own tech stack after the next strategy offsite, but before the Department of Commerce pulls the plug.

SpaceX Reaches $2 Trillion Market Valuation on First Day of Trading

SpaceX launched on the Nasdaq on Friday, breaking the $2 trillion mark. The stock jumped more than 19% to $160.95, peaking at $176.52, and more than 500 million shares changed hands on the first day—more than any other IPO this year. The opening price was $150, about 11% above the IPO price of $135, but it remained below the $175 that trading desks had anticipated. This makes Elon Musk the first documented trillionaire, and Goldman Sachs, as the lead-left bookrunner, gained over 2%. While SpaceX soared, other space stocks fell: Redwire down 11%, Rocket Lab down 10%, and the Procure Space ETF down 7%. COO Gwynne Shotwell, who herself was 'not sure if we would go public,' called the timing right. Anthropic and OpenAI have already confidentially filed their prospectuses. → www.cnbc.com

Synthszr Take: In May, we wrote that SpaceX was escaping the gravity of valuation models. Now it's $2.1 trillion, and the interesting number isn't in the stock price, but in the context: Redwire and Rocket Lab are imploding by double-digit percentages while a single name sucks up all the oxygen in the industry. This is the mechanism of an aggregator, applied to an industry that uses physical rockets. Anyone trading 500 million shares in a day isn't buying cash flow; they're buying a narrative about Starlink, Mars, and the question of who owns low Earth orbit. Westwood partner Dan Alpert calls the book 'pretty robust,' skeptics call it a valuation risk, and both are right. The real test won't come on opening day, but when the euphoric retail wave subsides and Starlink has to deliver what the stock price has already priced in. For Anthropic and OpenAI, the message is clear: the market is paying for vision today, and the window is open as long as the momentum lasts. .

OpenAI Acquires Kiel-Based Startup Ona

OpenAI is acquiring Ona, officially Gitpod GmbH, a provider for managing long-running AI agents. The purchase price was not disclosed. The core idea: Ona runs coding agents in cloud sandboxes that stay online even when the developer turns off their computer. This allows agents to work on tasks that stretch over several days instead of dying when the laptop goes into standby. It also includes security guardrails: Ona blocks suspicious programs via hashing, locks access to sensitive keys, and cuts outgoing connections to questionable servers. OpenAI is integrating the technology into Codex, which it claims has over 5 million weekly users. The entire Ona team will move to the Codex team. This is the first acquisition since the purchase of Promptfoo in March. → Techpresso

Synthszr Take: The real bottleneck for coding agents was never the model, but the runtime. An agent that stops when you close your MacBook is fine for autocomplete, but not for a three-day overnight refactoring job. Ona solves this, and OpenAI is buying the infrastructure layer that, in our code-crash vendor map, was previously occupied by Devin and the CI runners. This shifts Codex from an IDE helper to a continuously running background worker, and that's precisely where the market is being decided right now. The security story is interesting (sandbox deletion, hash-based blocking, credential protection) because it addresses the regulatory and compliance concerns that have held back large enterprises. The price: even deeper workspace lock-in, as the agents now run in OpenAI's cloud instead of on your machine. Anyone finalizing their coding stack in 2026 should consciously evaluate this shift of execution to the hyperscaler cloud and keep an open-source fallback like Cline with their own key, before customer retention becomes too expensive.

China's 618 Festival Becomes a Test Bed for Three AI Commerce Models

At the 618 Shopping Festival, China's largest mid-year retail event, three completely different AI commerce systems will compete for the same customers for the first time in early June 2026. On May 11, Alibaba fully connected its assistant Qwen with Taobao: users can search, compare, order, and pay via Alipay from over 4 billion items on Taobao and Tmall without leaving the Qwen flow. ByteDance took a phased approach, connecting its chatbot Doubao (345 million monthly active users in March) with Douyin's commerce platform, including a dedicated 'Help Me Choose' entry point. On June 8 and 9, Tencent opened WeChat's AI ecosystem to external developers; JD.com, Meituan, Didi, and Ctrip are the first to connect, with JD.com using an agent-to-agent integration. While Alibaba is building a direct bridge between the AI layer and its platform, and ByteDance is siloing its AI within its own commerce ecosystem, Tencent doesn't want to own the catalog at all. Instead, it aims to make WeChat the orchestration layer for partners who already have logistics and inventory. The WeChat integrations are still in development and testing; they are not yet available to end customers. → Hello China Tech

Synthszr Take: What we described in January as the 'new battle for e-commerce' is happening here in fast-forward, only now the tension is a product decision. Alibaba is running two parallel sales systems for the same goods: one optimized for user trust, and one that keeps the advertising cash flows alive. That's bold because no one cannibalizes a 70% profit base for fun. But the logic holds up: whoever captures the intent before the search box controls the interface, and in the digital world, value is created at the interface. Tencent is playing the most interesting card by not wanting to own the catalog or warehouse, but by making WeChat the orchestration layer on top of JD.com, Meituan, and Didi, which already have those capabilities. This is exactly the pattern WeChat demonstrated ten years ago with taxi services, only this time the service layer is an agent. The open question is not whether AI can recommend products, but who can convert intent into revenue without destroying the trust that directs it in the first place.

China Crosses AI App Threshold: 446 Million Users in a Quarter

As of March 2026, QuestMobile counts 446 million monthly active users of AI-native apps in China, an increase of 135 million since November 2025. This represents 43.4% growth within four months. ByteDance's Doubao leads with 345 million MAUs, followed by Alibaba's Qwen (166 million) and DeepSeek (127 million). Qwen jumped from 6th to 2nd place, adding 126 million users in the quarter alone, driven by the Hongbao promotions during the New Year festival. The average activity rate for the quarter is 33.5% for Doubao, 21% for DeepSeek, and 17.1% for Qwen. Per user, there are 87.1 sessions and 173 minutes of usage per month, each about a third more than in November. After the New Year peak, Doubao's DAUs dropped back to around 140 million, Qwen to 30 million, and Tencent's Yuanbao to 9 million. → Hello China Tech

Synthszr Take: Gaining 446 million active users in a single quarter is a distribution feat that no Western provider can match at this speed. But the interesting number isn't in the showcase, it's in the fine print: after the Hongbao rush, Doubao retains 140 million of its 345 million, while Qwen plummets from the top to 30 million DAUs. Red envelopes buy downloads, they don't buy habits. What really matters are the 173 minutes of usage time and the 33.5% activity rate for Doubao, because those show that a real use case is sticking, not just that a voucher was cashed in. In May, we wrote that China's formula for success depends less on chips and more on distribution and speed; this report is proof of that. Anyone looking purely at user numbers here is backing a losing horse. The second half will be decided by retention, and so far, only Doubao has proven it has staying power beyond the promotion.

WeChat Becomes an Agentic Everyday Companion

Tencent is pushing e-commerce deeper into the WeChat ecosystem, and this is the next step in a development we've been tracking for months. Alibaba has started wiring agent capabilities directly into Taobao, transforming a super-app from a pure service layer into an action layer that triggers transactions itself. Meituan is following suit, integrating models into real ordering processes, from food to local services. In China, 90% of online commerce already takes place on platforms; individual merchant websites play almost no role. Tmall alone brings together over 150,000 merchants and 200,000 brands. Now, platform operators are inserting generative artificial intelligence between the user and the product, capturing the interface a second time.

Synthszr Take: Back in 2017, in my book, I described WeChat as a transformational product par excellence because it simply absorbs and commoditizes mobility services like Uber. What's happening now is the logical continuation: the agent handles the order, you no longer tap through menus. Whoever controls the action layer controls the customer relationship, and that's precisely where the lever for monetization is. For European brands, this means: in China, you no longer build your own touchpoint; you become an ingredient in the experience cycle of Alibaba, Tencent, or Meituan. As we wrote at the end of May, falling chip prices and Tencent's push into AI agents are cushioning the impact of export controls—the pace remains brutal. The question is not whether the action layer is coming, but whose model will run it. Those who watch from the sidelines instead of building it themselves will later only be negotiating the commission.

Google Director: 'Management Has Lost Its Moral Compass'

René Mayrhofer, Director of Android Platform Security, has quit Google. In an internal farewell note titled 'Google Management Has Lost Its Moral Compass,' dated May 18, he writes that his resignation became 'unavoidable' after Google allowed the Pentagon to use its AI models for classified work. Mayrhofer also criticizes management for shelving CO2 neutrality goals due to the energy consumption of AI models. Business Insider obtained and verified the note. This isn't the first rift among the staff: in May, DeepMind employees formed a union to protest military use, and in March, OpenAI backtracked on its own Pentagon contract after internal protest. Mayrhofer wasn't a fringe activist; he was responsible for the security architecture of Android. → Business Insider

Synthszr Take: 'Don't be evil' hasn't been in the code of conduct since 2018, and Mayrhofer's departure shows how far the company has strayed since then. What's interesting isn't the outrage; we've seen that since the Maven era. What's interesting is who is leaving: the man responsible for the security of over three billion Android devices, the kind of person a corporation can't replace by posting three job openings. When the people with the most leverage leave the building, Google quietly loses its real moat: talent. In May, Pichai acknowledged the fears of Gen Z, while management was scrapping climate goals and prioritizing compute hunger above all else. The honest lesson for anyone building AI is this: you can't write a stance into your values charter and then remove it for the first billion-dollar contract without your best engineers noticing and leaving.

Enterprise AI Gaps Fuel 'Shadow AI'

While companies are still polishing their AI strategies, employees are already pulling their own tools out of their pockets. A study published Thursday by Wakefield Research and PagerDuty of 1,250 office workers shows that two-thirds use AI tools at work without explicit permission. 88% have fed work-related information into public chatbots, including customer data (34%) and confidential documents or financial data (31%). There were consequences: 53% were told to stop, and 48% received a formal warning. The reasons are revealing: 81% suspect that senior management plays by different AI rules than they do, and 72% are convinced they understand AI better than their own IT department. Tim Armandpour, CTO of PagerDuty, calls this a 'massive enterprise liability' and advises channeling that energy into vetted platforms with governance instead of slowing down adoption. → The Deep View

Synthszr Take: 86% of companies have an AI policy, yet two-thirds of employees are going rogue. This is the most honest market research on the state of enterprise AI I've seen in a long time. The finding isn't an IT security problem; it's a translation problem between tool logic and decision-making culture. When 48% are willing to risk a formal warning just to summarize meeting minutes, the organization has already lost control of a need it thinks it's managing. Bans don't create security here; they create shadows, and in the shadows, customer data ends up in a public model without any guardrails. By the way, the 81% who suspect a double standard are usually right: executives are experimenting with ChatGPT while compliance emails are sent to the rank and file. If you don't channel this energy into a sanctioned corridor with data protection and clear decision rights, you won't have a cautious workforce tomorrow morning—just one that's better at hiding.

Anthropic and TCS: Remote Instead of Forward Deployed Engineers

Anthropic has teamed up with Tata Consultancy Services, the Indian IT giant, to bring Claude to enterprises. TCS is building a dedicated business unit for this, getting early access to new model releases, and equipping its more than 50,000 employees with Claude. Together, they plan to build solutions for financial services, healthcare, telecommunications, and aviation; Diligenta, TCS's UK insurance subsidiary with over 22 million customers, plans to use Claude for customer service. The pattern isn't new: earlier this year, Anthropic partnered with Infosys, while OpenAI brought on Infosys and HCLTech. The timing is notable: investors are openly questioning the viability of India's $315 billion IT services business, with TCS stock down about 34% this year and Infosys down 31%. According to Anthropic, India is its second-largest market. → AI Secret

Synthszr Take: Here, Anthropic isn't buying a technology partner; it's buying a distribution channel with legs. TCS has what Anthropic lacks: access to the world's IT departments, established relationships, and people on the ground who can actually integrate a model into an insurance process. In May, we wrote that Anthropic had overtaken OpenAI in enterprise customers, with 75,000 in direct sales. This partnership is the next level: if your own sales team can't scale, you rent the 50,000 consultants of a corporation with its back against the wall. And that's exactly what's fascinating about the 34% stock decline. TCS feels the old manpower logic of IT outsourcing coming under pressure and is looking to the Claude code ecosystem for a way out of its own predicament. Whether the two will save each other or just extend the transition period depends on the edge case: rolling out a model is easy, but operating it cleanly in a healthcare process—where a technically correct answer can be emotionally catastrophic—is the real work.

Pokémon Go Mapped the World. Now That Mapping Navigates Military Drones

An AI model trained on data from Pokémon Go is set to help military drones navigate in areas without GPS. The augmented reality game, launched in 2016, garnered over 800 million downloads; a 2021 update introduced Pokéstops, where players scanned real-world locations with their cameras and uploaded them for in-game rewards. Niantic collected these location scans and used them to train foundation models that recognize and interpret physical spaces. In 2025, Niantic sold its gaming division for $3.5 billion to the Saudi-funded Scopely, while the spatial data remained with the spin-off, Niantic Spatial. In December, Niantic Spatial announced a partnership with Vantor, a provider of positioning software for drones, which in February secured a contract with the US Army worth up to $217 million. Both companies stress that the game scans were not directly passed on to Vantor, but only fed into Niantic's models. Privacy advocates like Tom Sulston of Digital Rights Watch point to the old pattern: people who play for free don't read pages of terms and conditions. → Techpresso

Synthszr Take: 800 million people mapped the world for free, believing they were catching Pikachu. This is the casual economy's business model in its purest form: you play, you provide, you are the product. We've seen this before with Strava, whose heatmaps revealed the locations of military bases; in March, we wrote about Meta's smart glasses that see everything. The common denominator isn't a company's malice, but the architecture: an opt-in checkbox from 2021 becomes the training data for GPS-independent navigation across three changes of ownership, and no one along the chain ever explained this to the player. The punchline is in the sale price: $3.5 billion for the games, but the real value remained with the spin-off holding the scans. Anyone building an app without a paywall today should ask what will happen to the collected data in five years and two sales later, because yesterday's terms and conditions cover every use case of tomorrow. Regulation based on the 'best interest of the user' standard would be the right lever, and that's a political decision to be made, not one to wait for after the next data scandal.