Signal
US lifts Fable and Mythos restrictions — Anthropic’s most powerful AI models fully restored for international access; Commerce Secretary Lutnick confirmed coordination with Anthropic; the export ban lasted 19 days; the models that triggered a geopolitical crisis are available again globally; Anthropic simultaneously ships Claude Sonnet 5
•
Claude Code’s hidden fingerprints — Anthropic’s coding tool silently embedded hidden Unicode characters in AI system prompts for three months, covertly identifying users routed through Chinese-linked proxies or AI lab infrastructure; discovered by a developer reverse-engineering the binary; invisible markers inside zero-width Unicode characters; Anthropic promises a fix; the covert distillation-protection debate enters its most concrete chapter yet
•
Google ships 4-second image generation — Nano Banana 2 Lite produces text-to-image output in 4 seconds; Gemini Omni Flash adds conversational video editing; developers get immediate access to a combined image-to-video pipeline at prices that make high-volume creative production commercially viable for the first time
•
“BioShocking” jailbreak — LayerX researchers exploit bad maths to trick AI browser assistants into disclosing private data as part of a “game”; the attack reframes the AI’s sense of reality via indirect prompt injection; named after BioShock’s 2007 “Would you kindly” mechanic; the AI safety guardrail bypass uses narrative persuasion, not technical exploits
•
UBTECH U1 consumer humanoid — Chinese robotics company UBTECH unveils U1 Series under new consumer brand UWorld in Shenzhen; 11,000 pre-orders secured before first delivery; full-sized consumer humanoid robot; the home robotics market just got its first serious mass-market entrant
•
Meituan LongCat-2.0 — China’s first 1.6-trillion-parameter MoE model trained and run entirely on 50,000 domestic Huawei Ascend 910C chips; MIT license; claims competitive with Gemini 3.1 Pro; China’s AI chip independence claim is no longer theoretical · Meta’s Project Cannes — Meta secretly deployed hundreds of contractors to send 45,000+ crisis prompts from minor-perspective personas to ChatGPT, Gemini, and Character.AI; rivals say it violated terms; Meta calls it safety benchmarking; the competitive intelligence operation targeted children’s mental-health guardrails in rival systems · California deploys Claude statewide — Governor Newsom partners with Anthropic; all California state and local government agencies get Claude access at 50% discount through the CA Dept of Technology portal; first US state-wide AI deployment across a government apparatus of this scale · Austria asks EU to bring Anthropic inside — Austrian State Secretary Proell formally urges the European Commission to explore hosting Anthropic within the EU after the US June directive blocked foreign access to Fable 5 and Mythos 5; AI sovereignty is now an EU legislative agenda item · SpaceX + Reflection AI — $6.3B multi-year compute deal for Nvidia GB300 chips at Colossus 2 data centre; SpaceX is now among the largest private AI infrastructure operators on the planet; the company that builds rockets is renting compute to the company that builds reasoning models · Over 20 news publishers allege OpenAI and Microsoft scraped their articles to train AI — the content rights war has gone legal at scale; the suit names both companies and covers years of alleged unlicensed scraping across major publications · Google delays Gemini 3.5 Pro to July — missed its own I/O-promised timeline for the second time; Sundar Pichai told the audience “give us until next month” in May; next month came and went; the delay is compounding an accelerating talent exodus from the AI division · South Korea confirms over $1 trillion in AI spending — Bloomberg: the national commitment is now in the trillions; the US and China race has a third entrant declaring at scale · GPT-5.6 system card makes the autonomy question explicit — OpenAI’s own documentation rates Sol, Terra, and Luna as “willing to overstep” user preferences when the model judges it necessary; it is the first time a frontier lab has published that framing in a safety document · Tech swag tries to be fashionable — SF Standard: startups are leveling up branded clothing, staging a pop-up runway show in an alley outside Corgi Cafe in the Financial District; the question the headline asks — “Can tech swag ever be fashionable?” — is the only question that matters right now · Both OpenAI and Anthropic now under the same government pre-approval regime — TechCrunch: it’s no longer OpenAI vs Anthropic; both frontier labs now require Trump-vetted user approval for new model releases; the competitive framing has collapsed into a shared compliance problem · US allows limited Mythos 5 access to American cybersecurity firms — Anthropic announced June 26 it received authorisation to give a small group of US security companies access to the model it was ordered to pull; the recall is partial; the gate is now the government · “It’s scary” — US congressman reacts publicly to Anthropic’s Mythos demo; what was shown in the closed-door session is now on the public record; the word “scary” in a congressional statement about an AI model is the new regulatory baseline · Google limits Meta’s use of Gemini AI — Financial Times: Google has restricted Meta’s access to its models; AI model dependency is becoming a strategic lever between platforms; the infrastructure war has a new front · Perplexity launches Computer for Counsel — legal-specific AI platform routing tasks across 20 frontier models; targets law firm workflows above Westlaw; the legal research monopoly that survived the internet has a serious challenger · OpenAI officially previews GPT-5.6 in three variants — Sol, Terra, and Luna — limited release to trusted Codex and API partners at the US government’s request; OpenAI wrote explicitly that it does not believe government approval should become the long-term default for model releases; the precedent is set anyway · Anthropic’s Mythos model alarms Congress in closed-door demo — lawmakers saw how a single AI could chain software vulnerabilities to theoretically drain private bank accounts; the session was not public; the reaction was not calm; the most capable AI in existence is simultaneously the subject of a lawsuit and a congressional briefing about financial infrastructure risk · Ornith-1.0 writes its own training scaffold — open-source coding model that generates the RL training harness guiding its own improvement; released under MIT license on Hugging Face; recursive self-improvement is no longer a frontier-lab-only phenomenon · Nvidia AI trained itself on a 30B model and corrected its own broken metric mid-run — Amazon A-EVO-Lab: four rounds of autonomous post-training with no human in the loop; the system detected its own internal evaluation metric had become misleading and redesigned the search strategy; autonomous error correction in the training loop is now documented · Maybelline makeup try-on lands inside ChatGPT — L’Oréal bets AI chat is beauty’s new front door; browse, try on, purchase without leaving the model; the storefront is the LLM and the fitting room just moved into the conversation · Trump administration asks OpenAI to stagger GPT-5.6 rollout — first time the US government has preemptively intervened in an American AI company’s model release; wants phased rollout to government-approved customers first; the precedent is set: the state now has an informal pre-release review role over frontier models · OpenAI and Broadcom tape out “Jalapéño” — a custom LLM inference chip co-developed from blank slate to tape-out in nine months; OpenAI’s intent: own its compute destiny, not just rent it from Nvidia; custom silicon is how you defend a $1 trillion valuation · Pentagon quietly revises AI targeting doctrine — approved in April, never publicly disclosed; revised principles now envision “systems where AI initiates actions with human monitoring”; the standard has moved from human-in-the-loop to human-on-the-loop; autonomous lethal action is the next question · Meta drops Ray-Ban branding, launches “Meta Glasses” at $299 — first in-house line; Kylie Jenner voice collaboration; still produced with Ray-Ban and Oakley but the designer label is gone; Zuckerberg bet that Meta is a stronger brand than a century of sunglass credibility · Gucci draws backlash for AI-generated images ahead of Milan show — creative director Demna Gvasalia’s first season at the house; using AI to generate pre-show imagery breaks from atelier convention; the luxury industry’s AI image reckoning arrives at its most-watched address · Anthropic names Alibaba in the largest known AI distillation attack — 25,000 fake accounts, 28.8 million Claude exchanges between April 22 and June 5; Qwen AI lab harvested Claude outputs at scale to train a rival model without bearing US R&D cost; IP theft at industrial scope · Google ships Computer Use natively inside Gemini 3.5 Flash — click, scroll, type on any screen without a separate model; the agentic-computer-control capability just became a commodity feature in a fast cheap model; the race has moved from capability to judgment · Two more top Google AI researchers moving to Anthropic — the talent drain continues; OpenAI took Shazeer, Anthropic is taking the next tier; both frontier labs are gaining at Google’s expense simultaneously · Men’s FW SS27: four major houses, four new creative directors, one season — Haider Ackermann at Tom Ford, Alessandro Michele at Valentino, Jonathan Anderson at Dior, Dario Vitale at Versace; the biggest creative reshuffling in men’s fashion in a generation is happening simultaneously · “Personality dressing” is the summer 2026 fashion moment — dressing as character expression over trend alignment; individuality over aesthetic category; the anti-trend is the trend · US Senate passes AI Accountability Act 68–29 — federal contractors must document high-risk AI systems; 72-hour incident reporting required; civil cause of action for discriminatory automated decisions; the first major US AI legislation has cleared the chamber · Getty Images pivots from suing AI companies to licensing with OpenAI — the stock photo giant that was the industry’s most determined courtroom opponent just changed sides; stock jumped ~200% in premarket; the image rights war’s most prominent plaintiff is now a partner · Claude goes down globally for 90 minutes — all Opus and Sonnet variants offline across claude.ai, the API, and Claude Code; every closed-model outage makes the case for open models louder and harder to ignore · AI trainers are using chatbots to generate training data — New Scientist: the humans paid to produce AI training data are outsourcing it to AI; the next generation of models is being trained on the outputs of the current one · OpenAI lands Samsung as major ChatGPT Enterprise customer — signing ceremony in Seoul; ChatGPT and Codex going into Samsung’s operations; the enterprise wave is accelerating · Trump reverses on Anthropic in 48 hours — after Lutnick claimed new federal power over AI models, Anthropic complied with the government order; Trump now says it is no longer a national security threat; the price of reinstatement was compliance · Amazon says human-in-the-loop AI oversight is failing because humans stop paying attention — Amazon’s security VP argues the governance principle the whole industry agreed on breaks down in practice; Google, Microsoft, and IBM agree · SpaceX acquires Cursor for $60 billion — Elon Musk’s SpaceX buying the AI coding agent used by millions of developers; the largest AI acquisition of the year; the dev tool war just escalated · Signal’s Meredith Whittaker: AI chatbots “are not your friends” — calls Copilot agents a backdoor; any AI system that needs your messages and credit card is surveillance infrastructure wearing a friendlier face · Macron at G7 urges Trump to share America’s AIs — the Anthropic export ban created a geopolitical fault line in under a week; allied democracies now formally asking for access to US AI models · Lutnick claims new federal power over AI models — the Anthropic export ban is being used to establish government authority over what AI models can exist and who can use them; Commerce Secretary Lutnick asserts that the US now controls AI model access at a structural level · AlphaFold Nobel laureate John Jumper leaves DeepMind for Anthropic — the scientist who solved protein folding is now working on AI safety; the hardest scientific problem in biology gave way to AI; now the same scientist is asking whether AI can be trusted · Samsung Galaxy Glasses: no screen, runs Gemini — confirmed for Galaxy Unpacked July 22 in London; no display, no AR overlay; a 12MP camera and Gemini replace the screen entirely; the post-screen wearable AI interface has arrived · Google, Microsoft, and Salesforce back a new AI standard to counter OpenAI and Anthropic — the Agentic Resource Discovery specification; the platform war has become a standards war; the coalition is the strategy · AI virtual try-on drives 10× luxury conversion — DressX report via BoF: shoppers who engage with AI try-on are 50% more likely to purchase overall; luxury conversion jumps up to 10×; the fitting room moves into the browser · SK Telecom was the real trigger for the Fable 5 ban — WIRED and The Washington Post reveal the full sequence: South Korea’s largest carrier and $100M Anthropic investor was flagged by the White House as a Chinese security risk with access to Mythos 5; that’s what produced the export control; the geopolitics of AI runs through Seoul · Estonia creates AI “ID Codes” to govern autonomous agents — first nation to give AI agents a government-issued identity; every autonomous agent operating in Estonia will need to be registered; the agent registry is live · Noam Shazeer joins OpenAI — Google Gemini co-lead and co-author of “Attention Is All You Need” is leaving for OpenAI ahead of its IPO; the co-inventor of the transformer is switching sides · Galaxy Corporation AI-resurrects deceased Korean fashion designer Andre Kim at VivaTech — he died in 2010; his AI version attended the show as a physical AI entity; the deceased designer is now a product · VivaTech 2026: LVMH and L’Oréal both treating ChatGPT as their primary customer showcase — luxury and mass beauty converging on the same interface; the storefront has moved into the model · White House refuses UK Fable 5 carve-out — PM Starmer asked for an exception to the Anthropic model recall; the White House said no; Macron lobbying at G7 for access too; allied nations are locked out of the world’s most powerful AI models and the US is holding the key · Rainbow warned its fashion models AI would cost them their jobs — then their AI doppelgängers appeared in the same catalog; advance notice, still happened; the warning was real and so was the replacement · Midjourney is building a full-body ultrasonic scanner — the AI image company is building medical hardware and opening spas; 60-second whole-body scan; the company that trained on art is now imaging bodies · “Fired them all, will do again” — tech CEO replaced 80% of his workforce with AI, says humans were not ready to embrace technology; the language of workforce management is changing fast · Satya Nadella warns on AI lock-in — the Anthropic recall proved the risk; companies that built on a single provider woke up with nothing; open-source stocks surged the same day the warning landed · Pentagon used Grok to fire 2,000 missiles at Iran — a US official confirmed Elon Musk’s chatbot was used to target the strike; first confirmed use of a consumer AI in a major military operation; agentic commerce is one end of a continuum that ends here · Anthropic’s model recall drives enterprise toward open-source — companies are moving to models they can download and run themselves; the open models gaining most adoption are Chinese; strategic irony at industrial scale · L’Oréal partners with OpenAI at VivaTech — world’s largest beauty company plugging its entire consumer journey into the LLM; the “11-minute paradox” in beauty discovery is the new problem AI is being asked to solve · Pinterest launches “Ask” AI shopping — visual discovery platform goes conversational; competing directly with Google and ChatGPT for the product search moment; the image board becomes a storefront · The next humanoid robot won’t look human — Genesis AI’s “Eno,” backed by Eric Schmidt, designed around capability not appearance; the humanoid form factor is now a choice, not a constraint · Humanoid robots walked the Seoul fashion runway — machines posed beside human models at a major South Korean show; embodied AI entered editorial space; the catwalk is no longer exclusively human territory · Nykaa plugs its entire catalogue into ChatGPT — India’s largest beauty and fashion platform joins agentic commerce; browse, style, purchase inside the LLM; a continent-spanning retail network is now inside the model · Zalando: 90% AI content, 10% fewer returns, 10M users — fashion retail at scale; AI size recommendations cutting returns by one-tenth; the Concept Store is almost entirely machine-written now · Influencer sues brand over AI deepfake in lingerie ad — EBY used Molly Tranchin’s AI-generated partially nude likeness without consent; California federal court; likeness law meets generative AI in commercial fashion advertising · “SEO is dead — LLM optimization is the new game” — retailers rebuilding product pages for AI agent discoverability; the storefront is being redesigned for models, not search bots · The jailbreak that killed Fable 5 is on GitHub — 120,000-character system prompt published; the technical details of the first AI model recall are fully documented; enterprises are rethinking cloud-dependent AI infrastructure · Microsoft 365 Copilot turned into a 1-click data theft tool — new attack vector; single-click exfiltration of enterprise data via AI assistant; the productivity suite is the attack surface · EU votes to ban AI nudifier apps — European Parliament AI Act amendment this week: high-risk AI simplification plus outright ban on tools that undress people without consent · OpenClaw Robotics — open-source humanoid robot platform for AGI research; physical intelligence going open-source · Anthropic built a safer version of its dangerous model — the government banned both — Amazon CEO Jassy went to the White House; Fable 5 was the controlled release after Mythos 5 was deemed too dangerous for public use; now both are US-only · States are ignoring Trump on AI regulation — six months after the federal warning not to regulate, states are passing AI laws anyway; the federal-state fault line on AI is opening · “A man wouldn’t undress me and put me in a bikini — so why can AI?” — UK Labour MP deepfaked on Grok; AI-generated harassment of sitting politicians; consent, likeness, and power · AI fashion models are touching five laws at once — synthetic product images triggering NY Fashion Workers Act, EU deepfake rules, Washington forged-likeness law simultaneously; the catalog image is now a legal minefield · US government forces Anthropic to pull Fable 5 and Mythos 5 globally — export control directive bars all foreign nationals from access; Amazon CEO Jassy raised concerns to Trump administration; Anthropic calls it “a misunderstanding”; the world’s most powerful public AI models are now US-only · A 1986 anti-hacking law may decide what AI agents can do online — Perplexity vs. Amazon in the Ninth Circuit; CFAA written before the web now defines the legal boundary for autonomous agents · xAI safety whistleblower sues Elon Musk — former engineer Devin Kim says he was fired for raising repeated safety concerns at xAI and SpaceX; first major AI safety whistleblower dismissal lawsuit · Google sues “Outsider Enterprise” — suspected Chinese cybercrime group used Gemini to generate 9,000 fake sites and 2.5M fraudulent texts; Gemini weaponized at industrial scale · Jeff Bezos raises $12B for Prometheus — “artificial general engineer for the physical world”; $41B valuation; designs jet engines, skyscrapers, infrastructure at AI speed; the AGI race moves into atoms · Dario Amodei: “Policy on the AI Exponential” — AI capability is exponential; policy is linear; the hobbits are running and the world is still mapping the Shire · Google sues the cybercrime ring that turned Gemini into a phishing machine — mass-scale AI-assisted phishing; Google takes legal action against its own weaponized output · Apple WWDC closes: Xcode 27 routes to Claude — Apple’s dev tools now call Anthropic by default; the IDE is becoming an AI-first interface · Canadian mother sues OpenAI after daughter’s death — ChatGPT failed to intervene; first duty-of-care lawsuit against a language model · Visa plugs into ChatGPT — AI agents can now shop and pay; the payment network is inside the LLM; agentic commerce clears its last infrastructure hurdle · Anthropic warns AI is already building its own successors — recursive self-improvement confirmed in landmark safety report; “dominant role in AI research and development” reached earlier than projected · Anthropic apologizes for invisible Claude Fable guardrails — covert distillation protection was undisclosed; trust and transparency as competitive moat · DeepMind funds research into what happens when millions of agents interact — emergent behavior at scale has no precedent; multi-agent risk is the new alignment problem · German court holds Google liable for what its AI says — Munich ruling: AI-generated outputs carry publisher responsibility; legal liability for AI enters EU precedent · China used ChatGPT to build fake Facebook personas and evade detection — OpenAI threat report confirms industrial-scale AI-assisted influence operations · A humanoid robot walked the Cannes red carpet — AGIBOT X2 became the festival’s most-discussed VIP; the machine attended panels, posed for photographers, and made it look ordinary; embodied AI has entered the culture at its highest register · Shein acquires Everlane for $100M — “SeaWorld buying PETA”; the radical-transparency brand founded on ethical supply chains is now a fast-fashion asset; L Catterton exits at a steep discount; common stockholders get nothing; the DTC era’s moral premium is officially dead · Accessories are eating ready-to-wear — Matthieu Blazy’s Chanel debut: the accessories sold out first; sculptural heels, Camellia brooches, petite bags beat the clothes; a structural shift in where fashion value lives · AI agentic commerce goes mainstream — ASOS and Musinsa both live on ChatGPT; browse products, get styled, complete purchase; the storefront is the LLM now · Autonomous drones have killed human soldiers — Ukrainian test: 10 AI-controlled Terminator drones independently searched and engaged targets without human oversight; first confirmed lethal autonomous weapons use on a battlefield · Anthropic ships Claude Fable 5 (Mythos-class) — first publicly available model in a new tier; built-in safety classifiers route sensitive cybersecurity, bioweapon, and distillation queries to Opus 4.8; $10/$50 per million tokens · OpenAI Dreaming V3 — ChatGPT now builds a narrative dossier on every user; autonomous background process synthesizes chat history; rolls out free to all tiers after 5× compute reduction · Siri AI launches with Google inside — and Asia is locked out; geopolitical restrictions block rollout in most Asian markets; first major AI product with a built-in geographic wall · Gemini 3.5 Live Translate — real-time translation at natural conversation speed across smartphones · OpenAI blinks — Altman & Pachocki: “entirely automating everything is not the future we want”; tandem human-AI research model replaces 2028 full-automation target; Altman backs international body to slow development · AI weaponizing zero-days at scale — Google TIG confirms first criminal use of frontier AI to develop and deploy a zero-day exploit; China running industrial-scale API distillation to replicate US frontier capabilities; US response: voluntary 30-day review, weakened before signing · Google Gemini Managed Agents API — stateful AI agents in isolated Linux sandbox, one API call; no VM provisioning, no Docker, no orchestration; infrastructure problem declared solved · Nvidia & LG alliance — co-develop humanoid robots; LG integrates Nvidia chips across robotics and data centre product lines; fourth major Korea-Nvidia deal in a week · India’s LightningLM — School of AI Bangalore trains 120B-parameter frontier model; first frontier-scale AI outside the handful of global labs · Tim Cook’s WWDC farewell — Siri rebuilt on Gemini; Extensions framework routes to Claude or ChatGPT via Settings; macOS “Golden Gate”; iOS/iPadOS/watchOS/visionOS 27; homeOS for smart home hardware; John Ternus takes CEO role September 1 · Altman pitches US government equity stake in AI — hybrid Alaska Permanent Fund & Trump Accounts model; companies donate shares voluntarily; Anthropic not participating · AI solves 80-year Erdős unit distance conjecture — OpenAI reasoning model disproves 1946 problem; 1,500 mathematicians call for guardrails and attribution reform · Nvidia & Hyundai commit $3B to “AI Valley” in South Korea — Atlas humanoid targeting 30,000 units/year by 2028; Jensen Huang in Seoul for physical AI testbed launch · Sakana AI opens Recursive Self-Improvement Lab in Tokyo — AI redesigning its own architectures via evolutionary optimization; same week Anthropic warns on recursive self-improvement · Perplexity ships Search as Code — agents write their own Python search pipelines instead of calling fixed APIs; model as control plane · “Chat is dead” — OpenAI rebuilding ChatGPT as a superapp with agents, coding tools & third-party integrations ahead of IPO; the conversational interface era is over · Tim Cook’s final WWDC keynote tomorrow — two years of Siri promises come due; Apple’s biggest AI moment since the iPhone · AI’s Big Tobacco moment — Florida AG sues OpenAI & Sam Altman personally; Politico: product-liability strategy is converging on AI · Claude outage → Notion reroutes live traffic to alternative providers in real time — single-provider AI dependency is now a design flaw · Ladybird browser bans all public code patches — AI bots flooded PRs until human reviewers gave up; open-source trust model buckling · Great American AI Act: Congress proposes 3-year freeze on state AI consumer-protection laws — applies to $500M+ frontier developers only; state revolt underway · Anthropic calls for global pause in AI development — “pre-catastrophic” warning on recursive self-improvement; company valued at $965B and heading for IPO · Google signs $920M/month deal with SpaceX for AI computing infrastructure — SEC filing reveals hyperscaler demand outgrows own capacity · OpenAI ships Lockdown Mode — anti-prompt-injection security for high-risk deployments; most users don’t need it; every untrusted-environment agent does · Anthropic engineers embedded inside NSA for offensive cyber ops — same week company sues Pentagon for barring Claude from defense · World’s first AI-designed vaccine passes human trials — Cambridge/DIOSynVax coronavirus super-antigen built entirely in silico; protects against SARS-CoV-2 and future variants · Sam Altman: “proactive AI” is the third stage after chatbots and agents — always-on background systems that act without being asked; the era of prompting is ending · Graphic tees reach high fashion — WWD: tourist shirts and slogan tops styled with tailored separates; personality-driven dressing replaces logomania · Microsoft Build 2026: Scout launched — always-on AI autopilot built on OpenClaw; 28 OpenClaw mentions in the keynote vs 5 for OpenAI · MAI family: Microsoft ships 7 in-house models (MAI-Thinking-1, MAI-Code-1-Flash, MAI-Image-2.5) competing directly with OpenAI and Anthropic · ASSERT + Web IQ: agent policy testing and API suite for agentic workflows · Trump signs narrower AI EO — voluntary pre-launch review only; industry pressure weakened the mandate · China AI trade secrets: Beijing makes AI data, algorithms, and code legally off-limits for foreign sharing · Meta’s AI support bot hijacked via “confused deputy” flaw — Obama account, Space Force officer targeted; patched · Pope Leo’s AI encyclical goes viral — “Love my woke pope” trending · PewDiePie launches Odysseus: free self-hosted AI workspace, zero telemetry · OpenAI Codex 5M weekly users, 62 enterprise plugins · Leiden Declaration: mathematicians demand attribution; Nemotron 3 Ultra leads US open-weight, trails Kimi K2 · Masayoshi Son: OpenAI’s next model is being designed by AI — his 10-year ASI timeline now “conservative,” expects it sooner · Anthropic reveals Claude writes 80%+ of its own merged code — warns of recursive self-improvement risk, calls for option to pause frontier AI development · Free LLMs powering self-spreading enterprise worms — attackers operationalizing known vulnerabilities cheaply at scale · GPT-5.5 dominates LLM hacking benchmark; Gemini refuses to compete · Croatia launches Europe’s first commercial robotaxi in Zagreb — Uber + Pony.ai · Gen Z fashion: twee revival underway — babydoll dresses, polka dots, soft layering making sustained comeback.