inklede.

The dollar traded near a three-month low against the euro on Friday as concerns grew over the U.S. Treasury's plan to expand buybacks of longer-dated government debt. Treasury Secretary Scott Bessent said he may increase repurchases further, a day after the department pledged to at least double the size of its buybacks. The 30-year yield hit its highest level since 2007 this week. The dollar index fell 0.05% to 98.79, the euro rose to $1.1682 after touching $1.1711, and sterling reached $1.3675, its highest since February 11. Bitcoin surged 6.06% to $77,060. Fed funds futures price 35% odds of a September rate hike, rising to 69% by December, with Fed Chair Kevin Warsh set to speak at Jackson Hole next week.

UK AI infrastructure startup Callosum has raised $100 million in a seed round led by Atomico, with participation from Plural, DCVC, Sovereign AI and angel investors, following a $10.25 million round in February when it emerged from stealth. Founded by Cambridge neuroscientists, Callosum builds software that lets AI workloads run across different chips and models rather than relying on a single large model on identical hardware. It has partnered with chipmaker Cerebras and won praise from UK AI minister Kanishka Narayan, who said integrating Cerebras into Callosum's platform enables ultra-low-latency inference for previously impractical AI systems.

This month's Crunchbase roundup of overlooked startup deals is dominated by AI applied to physical-world problems. London's Greyparrot raised £20.3 million ($27 million) Series B to expand AI camera systems that identify materials on recycling conveyor belts, with customers including Waste Management, Veolia, Unilever and L'Oréal. Australian robotics firm Emesent secured $17 million (a $10 million equity round plus $7 million venture debt) to scale Hovermap, its GPS-free navigation tech for mines and tunnels, now deployed at 200+ sites. San Francisco's SoundHealth raised a $12.25 million Series A for FDA-cleared acoustic devices treating congestion. Nashville's Forge Industries raised $3.85 million to convert unrecyclable plastic into industrial fuel. New York's Cascade raised $3.5 million seed, backed by a16z's speedrun accelerator, to predict construction bids using public filings.

Plymouth-based ocean robotics startup Oshen has raised €4.27 million ($5 million) in a round led by Lunar Ventures, with AlbionVC, Twin Track, and Concept Ventures also participating, and is tripling manufacturing of its C-Star autonomous ocean sensors. Founded in 2022 by Imperial College engineers Anahita Laverack and Ciaran Dowds, Oshen builds 1.2-metre wind- and solar-powered robots that operate in constellations for months, transmitting data via satellite. In September 2025 three C-Stars penetrated the eyewall of Category 5 Hurricane Humberto under a NOAA contract, feeding data into the National Hurricane Center's forecast. Customers include the US Navy, Royal Navy, NOAA, the Met Office, and Météo-France. Production has scaled from 15 robots over three years to 15 every six weeks, with 100 more on order and a 50-strong ARIA-funded constellation planned for the North Atlantic.

Samsung Electronics announced it expects shareholder returns of between 90 trillion and 110 trillion won ($65.1 billion to $79.52 billion) in 2026, which it called the largest such program ever by a Korean company. The move follows SK Hynix's announcement days earlier of a 40 trillion won share buyback, marking a blockbuster week for shareholder returns among South Korea's chip giants. Samsung, whose stock is up around 135% year-to-date, said it will pay roughly 30 trillion won in cash dividends in the third quarter, including its regular payout, with further details finalized at a board meeting in late October. The company has been racing to catch SK Hynix in high-bandwidth memory chips used in AI systems, and will decide the remaining return details at a January 2027 board meeting.

Hyundai Motor is considering expanding production capacity at its Georgia Metaplant from 500,000 vehicles annually to between 700,000 and 800,000 units by 2028, CEO José Muñoz told CNBC. At the top end, the plant would surpass Tesla and Toyota to become the largest U.S. vehicle assembly facility by capacity. Muñoz said Trump's tariffs, including a 15% levy on South Korean imports, accelerated Hyundai's localization plans, which were already underway. Any expansion would draw on Hyundai's existing $26 billion U.S. investment commitment through 2028. The $7.6 billion plant currently builds the Ioniq 5, Ioniq 9, and Kia Sportage hybrid. Hyundai aims for 80% domestic production share of U.S. sales by 2030, up from roughly 40% in 2024, and is also evaluating a separate site for trucks and SUVs.

Former JPMorgan Chase executive Matt Zames is joining the Trump administration as an unpaid advisor to the Social Security Administration, CNBC has learned. He will help Frank Bisignano, his former JPMorgan colleague and now SSA Commissioner, modernize the agency's decades-old technology systems. Zames starts Monday at SSA headquarters in Baltimore. He rose to prominence at JPMorgan helping clean up the $6 billion

With EU AI Act enforcement powers active since August 2, 2026, EU-Startups profiles ten European regulatory-technology startups positioned to benefit. Highlights include Barcelona's Biorce (clinical-trials AI, €43.8M Series A in February), Paris's Cleo Labs (€1.5M raised in April, monitoring 25,000+ regulatory authorities across 106 countries), Berlin's Cortea (audit AI, €12M seed in June), Stockholm's Hybridity (DORA/NIS2/GDPR compliance, €2M raised in February, ~€5M total), Barcelona's NeuralTrust (AI agent security, €17.2M seed, described as the largest EU cybersecurity seed round at the time), and London's Outpost (cross-border tax/compliance infrastructure, €15M raised in March). Others include CertHub, Fortiv, Rippletide, and Rulemapping Group.

France remained a leading European tech funding market in H1 2026, with capital concentrated in AI, space, fintech and healthcare. Top raises included Eutelsat's €975 million debt financing for 340 new LEO satellites, Advanced Machine Intelligence's $1 billion seed round (co-founded by Yann LeCun and Alexandre LeBrun) for world-model AI research, and Mistral AI's $830 million debt raise to buy 13,800 Nvidia GPUs for its first French data centre. Other notable rounds: Alan raised €580 million across two tranches valuing it at €5.5 billion, Aura Aero raised €340 million for electric aircraft, Harmattan AI raised $200 million for defense autonomy systems, Pennylane raised $200 million, Morpho raised $175 million, Quobly raised €115 million for silicon quantum computers, and FineHeart raised €83 million for cardiac implant technology.

US stock futures traded mixed Thursday morning after Treasury Secretary Scott Bessent unexpectedly moved to increase long-term bond buybacks, doubling the cap from $2 billion to at least $4 billion starting September 9, ahead of the normal November review. The 30-year yield fell 9 basis points to 5.19%, its biggest one-day drop in over three weeks, and the dollar index dropped 0.75%. Bitcoin surged past $70,000 for the first time since June, triggering roughly $1 billion in short-position liquidations, the largest squeeze since 2021. The national debt surpassed $40 trillion. Separately, Walmart beat earnings estimates with $187.9 billion in revenue but its stock fell 6% after US same-store sales grew only 2.6%, below the 3.7% forecast. President Trump also threatened an "ECONOMIC D-DAY" against Iran.

Goldman Sachs projects roughly $7.6 trillion in cumulative AI capital spending from 2026 through 2031, while OpenAI and Anthropic's combined annualized revenue topped $105 billion by August 2026. Analysts say the industry needs annual recurring revenue past $1 trillion by 2030 to avoid write-downs on hardware with three-to-five-year lifecycles. Nvidia, whose gross margin sits at 74.9%, carries the most exposure if hyperscalers shift routine workloads to their own chips like Google's TPUs or Microsoft's Maia. Bulls counter that inference costs have fallen roughly 50-fold, from about $20 to $0.40 per million tokens, potentially expanding demand. Microsoft's AI business exceeded a $37 billion annual run rate in the March quarter, up 123% year over year, with Copilot passing 30 million paid seats by June.

Danish healthtech startup Aisel Health has closed a €1.7 million pre-seed round led by Caesar Ventures, with Nordic Web Ventures, LifeX, and Angel Invest joining alongside existing investors Rockstart and EIFO. Aisel is building an operating system for psychiatry that goes beyond AI scribes, aiming to surface organized patient context (from interview recordings, referrals, and prior notes) to clinicians before, during, and after consultations rather than simply generating more documentation. Co-founder Christian Houen says psychiatrists spend only about 40% of their time with patients, with the rest consumed by gathering and writing documentation. The platform is piloting AI interviews with patients' relatives, schools, and other parties for cases like ADHD diagnosis. Funding will support hiring and a UK market entry ahead of a planned seed round.

Uber and drone delivery company Zipline announced a partnership on Monday, August 17, to bring autonomous drone delivery to Uber Eats customers across the United States, with Uber also making a strategic investment in Zipline whose financial terms were not disclosed. The companies aim to reach 1 million drone deliveries per day by the end of 2029, with first deliveries beginning later this year in Zipline's existing US markets before expanding to dozens more cities. Zipline, valued at $7.6 billion after an $800 million raise earlier this year, has made 2 million deliveries across four continents and counts Walmart, Panera, Chipotle and Wendy's among its partners. The deal mirrors Uber's earlier partnership and investment in drone firm Flytrex.

IREN stock jumped 6% to $45.26 Friday after Microsoft formally accepted Horizon 1, the first of four data centers IREN is building for the company at its Childress, Texas campus, triggering the first billable revenue under a five-year, $9.7 billion contract. Shares gave back the gain within the first hour, closing near $42.54. Separately, Nvidia certified the Childress site with Exemplar Cloud status after testing its GB300 systems there, though Microsoft and Nvidia stock barely moved. TeraWulf rose 1% to $16.61 while Cipher Digital fell 5% to $16.35. IREN has secured a $3.65 billion investment-grade debt package tied to the Microsoft contract, plus a separate five-year, $3.4 billion cloud deal with Nvidia. More than a quarter of IREN's shares remain sold short.

Vienna-based spacecraft propulsion maker ENPULSION has agreed to acquire 100% of UK space propulsion company Lift Me Off, whose deal terms weren't disclosed and remains subject to UK government approval under the National Security and Investment Act 2021. ENPULSION, which builds electric propulsion systems using field-emission electric propulsion (FEEP) technology and has more than 320 systems currently in orbit, wants to combine LMO's chemical propulsion and propellant tank expertise with its own electric propulsion and Cortex integration layer to offer a complete spacecraft mobility architecture. LMO, founded in 2018 by Michel Poucet and Marcos Perez and based in Reading, has worked on European Space Agency missions. The acquisition follows ENPULSION's €22.5 million growth funding round led by Nordwind Growth earlier this year.

Pixelgen Technologies, a Stockholm-based life sciences company, has closed an oversubscribed €13.26 million ($15.5 million) Series B round led by new investor Flat Capital, with continued backing from Industrifonden and Navigare Ventures. Flat Capital CEO Rickard El Tarzi and life sciences veteran Ramon Felciano are joining Pixelgen's board. Founded in 2020 by CEO Simon Fredriksson and three co-founders, the company makes the Proxiome Kit, which it says is the first platform for high-throughput mapping of molecular cell architecture at nanoscale resolution, useful in immunology, haematology, oncology and cell therapy research. Pixelgen previously raised a €5.13 million seed round in 2022, a €6.24 million Series A in 2023, and a €12.5 million EIC Accelerator grant and equity package in June 2025. The new capital will fund global commercialization.

Ali Hussain, founder of New York fintech startup Tabs, built a $400 million company despite having no computer science background. Hussain earned a humanities degree at Cornell, won a Marshall Scholarship to Oxford but left his PhD after two months, worked at Boston Consulting Group, then spent six years as an early operations hire at startup Latch. In 2023 he founded Tabs, an AI platform automating revenue recognition, billing and collections, partnering with technical co-founder Deepak Bapat. Tabs raised a $4 million pre-seed round led by Primary Venture Partners and One Way Ventures, has since raised roughly $92 million total, employs about 180 people, and has grown revenue at triple- to quadruple-digit year-over-year rates.

German agritech company Computomics has raised €6.3 million in Series B funding to scale its climate-smart breeding platform, which combines genomic data with environmental measurements to help crop breeders predict variety performance under changing conditions. Convent Capital Agri Food Fund led the round with €5 million, joined by existing investors High-Tech Gründerfonds, MBG Baden-Württemberg and Amathaon Capital, plus support from the EU's InvestEU Fund. Founded in 2012 in Tübingen, Computomics applies machine learning to genomic and field-trial data through its ×SeedScore platform, helping breeders identify candidates suited to hotter, drier conditions. CEO Sebastian J. Schultheiss said the technology lets breeders assess climate resilience earlier, addressing breeding cycles that otherwise select varieties for climate conditions years out of date. The funding will expand deployment across more commercial breeding programs.

UK edtech startup Medly AI has raised $8 million in seed funding led by Felix Capital, with participation from Eka Ventures, Ada Ventures, and several angel investors including Andrey Dobrynin of InvestEngine. Founded by UCL medicine graduates Dr Paul Jung and Dr Kavi Samra, Medly built an AI tutor using seven large language models combined with neuroscience and teaching-theory approaches, offering personalized exam prep with handwriting recognition for subjects like math and chemistry. Since launching in February 2025, the platform has reached over 400,000 UK users averaging about an hour of daily use. The company is running a trial with more than 1,000 students on attainment impact and has begun international expansion, launching SAT prep in the US with AP and ACT support planned by year end.

The Idaho National Laboratory, a Department of Energy facility, is investigating whether Chinese-made lidar sensors pose security risks if widely adopted in US vehicles, TechCrunch has learned from two people familiar with the effort. The review is funded by an unnamed company or group of companies in the EV and autonomous vehicle industry. Rivian, GM, Ford, Kodiak, Lucid, Nuro and Uber said they were unaware of it; Nvidia, Zoox and Aurora did not respond. The probe comes as lawmakers including Senator Tammy Baldwin and Representative John Moolenaar push bills to restrict Chinese lidar, citing espionage risk. Innoviz CEO Omer Keilaf outlined cybersecurity concerns including mass sensor disabling and data exfiltration via compressed lidar feeds. Chinese suppliers Hesai and RoboSense, already on the Pentagon's 1260H list, did not respond to requests for comment.

A detailed technical deep dive traces what happens when a GPU executes a single global memory load instruction (LDG.E) on an RTX 4090, following the path from a vector-add CUDA kernel's compiled SASS instructions through registers, the load/store unit, the coalescer, the L1 cache, address translation (TLB), and onward toward the L2 cache. The piece measures real latencies through timing experiments rather than relying on undocumented specs: an L1 hit costs about 15.4 ns (40 cycles), a TLB miss costs about 4.4 ns (11 cycles), with further stages (L2, DRAM) still to be covered. It documents the L1 cache's 4-way set associative structure, 128-byte line organization, and virtual addressing, building a rigorous, empirically grounded picture of GPU memory hierarchy behavior.

Starcloud, a startup building satellites for orbital AI inference, has raised a $250 million extension to its March Series A, valuing the company at $2.3 billion. The round, led by Manhattan West Ventures, included Nvidia (about $25 million) and Cisco, plus Benchmark, EQT, Soma, NFX, 776, and others. CEO Philip Johnston says the money will fund a larger manufacturing facility and development of Starcloud-3, meant to fly on SpaceX's Starship. With Falcon 9 scheduled to end in 2028 and Starship still unproven, Johnston is racing to secure launch capacity. Starcloud plans to launch two 8kw Starcloud-2 satellites in 2027 for government customers, and is the only known operator of an Nvidia H100 GPU in orbit, feeding data to Nvidia's in-development Vera Rubin Space-1 space chip, targeted for a 2028 flight.

Nvidia research published Friday argues that the software harness wrapped around an AI model matters more than the model itself for long-horizon agentic tasks. Using a custom harness called Agentic Variation Operators, which includes a supervisor component that redirects a stuck agent, researchers got Claude Opus 5 to score 100% on the ARC-AGI-3 benchmark, a set of instruction-free 2D games. Without the harness, Opus 5 scored just 30%, still the best raw result among tested models. OpenAI, whose models scored under 10% on the same benchmark, separately found tripling scores possible by tweaking harness settings, but never reached 100%. Nvidia VP Adel El Hallack said the harness, covering memory, tools, and runtime, is what actually turns a model into an agent, and Databricks has similarly found harness choice can double AI costs regardless of model.

Cloudflare has published new research revisiting its 2021 defenses against remote Spectre attacks on Cloudflare Workers, its serverless platform running untrusted JavaScript for tens of thousands of tenants. Researchers built a proof-of-concept exploit in Cloudflare's production environment and found a limitation in Dynamic Process Isolation, the defense shipped in 2021, allowing a remote Spectre attack to leak up to 12 bits per second with 99% accuracy despite frozen timers and no shared memory. The attack used speculative type-confusion gadgets and cache-timing amplification via a remote WebSocket timer. Cloudflare says the flaw is already mitigated through an upgraded DyPrIs, the V8 Sandbox, and in-process isolation, and found no evidence of exploitation over three years. The paper is co-authored with researchers including Martin Schwarzl.

Apple has overhauled its EU App Store fee structure to resolve a long-running dispute with EU regulators over Digital Markets Act compliance. The company will replace its per-install Core Technology Fee with a flat 5% commission on digital goods sold in apps distributed outside the App Store or on the web. In-app purchase fees drop from 30% to 26%, though most developers still qualify for a discounted 15% rate through programs like the App Store Small Business Program. Alternative payment processing will cost 20% (or 10% for qualifying developers). Developers must commit to their chosen payment model for 12 months. Apple also loosened requirements for launching alternative app marketplaces, dropping the prior mandate that developers prove two years' tenure and one million EU installs, and added new ways to show financial backing, including public company status and VC funding.

Anthropic announced it is expanding access to its most capable model, Claude Mythos 5, for cybersecurity defense. The company is integrating Mythos 5 into Claude Security, now available to Claude Enterprise customers for scanning codebases and suggesting patches (with human review required), and is working with cybersecurity partners to embed Mythos 5 into their existing tools without giving end users direct model access. Anthropic is also launching a $35 million Defender Advantage Fund (0xDAF) in Claude credits to help patch and secure open-source software, building on $4 million in prior direct donations through its Project Glasswing initiative. It is expanding its Cyber Verification Program, which currently gives vetted defenders reduced safeguards on Opus and Sonnet models, to eventually include Mythos-class access, and continuing government partnerships focused on critical infrastructure protection.

TikTok and ByteDance have agreed to pay $400 million to settle a DOJ lawsuit alleging violations of the Children's Online Privacy Protection Act (COPPA). The suit, filed in 2024 under the Biden administration, claimed TikTok let millions of children under 13 use the platform and collected their personal data without parental consent, despite a 2019 settlement over similar allegations by its predecessor Musical.ly, which cost $5.7 million. The new agreement adds age-verification controls and parental oversight tools but does not require an admission of wrongdoing. The settlement lands days after Bloomberg reported TikTok disabled a harmful-content safeguard for 10% of U.S. users as an experiment, prompting a bipartisan letter from senators Marsha Blackburn and Richard Blumenthal to TikTok's leadership questioning the decision.

Google's Threat Intelligence Group is tracking three suspected Russian cyber espionage clusters, UNC6293, UNC7005, and UNC5976, that abuse legitimate authentication flows to target academics, aerospace and defense personnel, government officials, and think tank staff across Europe and the US. UNC6293, assessed as linked to the group formerly known as APT29, has run app-password phishing campaigns since 2025, impersonating the US State Department and more recently shifting to OAuth phishing that harvests verification codes. UNC7005, identified in February 2026 and tied to hospitality captive-portal redirects reported by Reliaquest and Microsoft, conducts app-password and device-code phishing against Microsoft and WhatsApp accounts, using lures like fake conference invitations (including a spoofed GLOBSEC forum registration) and anti-detection scripts, and can hijack linked WhatsApp devices to record calls.

Apple has laid off roughly 100 employees from its Vision Products Group, largely shutting down the Vision Pro gaming team and paring back the immersive video unit, according to Bloomberg. Another 100 people were cut from Siri and software groups. Bloomberg reports Apple's immersive videos, which can cost millions of dollars to produce, haven't been cost-effective given low active usage of the device, and Apple will now rely more on third-party 3D content. On the Siri side, Apple is creating new roles focused on Siri AI as it prioritizes AI features. Affected employees can apply for other positions. Apple says Vision Pro and visionOS are not being discontinued, and it continues developing AI smart glasses expected in 2027 that will not support 3D gaming or video.

Stripe and Ramp both moved into AI model routing within 70 minutes of each other on Wednesday: Stripe acquired OpenRouter (Bloomberg puts the price above $7 billion, Axios says over $8 billion in cash and stock), while Ramp launched its own internal router claiming roughly 40% lower cost for equivalent output. OpenRouter serves more than 400 models from over 80 providers and processes more than 10 trillion tokens daily. The piece argues routing, not model choice, is now the strategic layer: Stripe is targeting developers the way it already routes payment methods, while Ramp is targeting finance teams via its existing spend-management relationships. It cites examples of routing risk, including a support-agent case where a cost-cutting classifier misrouted fraud queries to a cheaper model, causing customer churn.

The FTC has opened a public comment period on a proposed policy statement targeting personalized pricing, where businesses use personal data to set the highest price an individual might accept. The agency lacks authority to ban the practice outright but says businesses that fail to disclose data-driven pricing may violate the FTC Act. Chair Andrew Ferguson said new industries are increasingly using this tactic without customers' knowledge. Examples cited include a food delivery service raising prices when data shows a customer can't leave home, a grocery chain charging families more for milk, and rideshare apps charging more if no competing apps are installed. Public comments split: most support strict limits, calling the practice discriminatory, while critics like commenters Deymond Lashley and Jessie Shettleroe warn that vague rules could eliminate loyalty discounts and coupons or simply legitimize the practice through disclosure requirements rather than banning it.

A new analysis using Artificial Analysis's Intelligence Index and cost-per-task data finds that the price of a given level of AI model capability has dropped roughly 56x in under six months, falling from $1.22 per task in February to $0.022 today. At that pace, a 100x price drop for a fixed capability level takes about a year, and the rate of decline is accelerating: the June-to-current frontier replaced fifteen of sixteen prior frontier models. The author illustrates this with a real research task, running roughly 10,000 papers through a model to check dataset reuse, which cost just over $100 today versus several thousand dollars in March, and was unavailable at any price a year ago.

Apollo Global Management, a private equity firm managing $938 billion in assets, has confirmed a data breach in a letter filed with California's attorney general. Apollo's HR chief Matthew Breitfelder said hackers used a social engineering attack to access the company's cloud environment between July 6 and July 10, stealing names, birth dates, home addresses, and Social Security numbers. It's unclear whose data was taken, employees or people at portfolio companies, and Apollo declined to say whether it paid a ransom.

The breach follows a Google warning about hackers, tracked under names including Falcon, Helix, Pink, and Redact, targeting private equity and financial firms including Blackstone, Bridgewater, and Bain Capital using IT-helpdesk impersonation scams. Google says some victims paid ransoms as high as $750,000.

Security researcher Jeremiah Fowler found that people-search service ClarityCheck left more than 9 million images, roughly 450 GB, exposed in an unsecured Amazon S3 bucket, despite marketing its reverse-image search as private and secure, Wired reports. The images, stored in folders labeled

Meta goes to trial Tuesday, August 18, in federal court in Oakland, facing claims from dozens of states that it designed addictive features, misled the public about safety, and illegally collected data from children under 13 in violation of COPPA. States including California, Colorado, Kentucky and New Jersey are pursuing the consumer protection claims before Judge Yvonne Gonzalez Rogers, while 29 states total are involved in the COPPA claims. Meta says states are seeking penalties up to $1.4 trillion, a figure the judge has called unreasonable. The eight-member jury serves in an advisory role only, with the judge deciding the final verdict. The trial is expected to last six weeks and may include testimony from Mark Zuckerberg and Adam Mosseri. Meta spent $2.4 billion on legal costs in Q2 2026 alone.

Anthropic launched Browser Use, a tool announced Thursday that gives Claude a structured, text-based view of a web page via its accessibility tree instead of relying only on screenshots and pixel coordinates. Developers access it through the Claude API using browser_toolset_20260801, alongside the newly generally available Computer Use, Skills API and Files API. Instead of clicking coordinates, Claude can reference page elements like ref_3 and act on them; the tool also lets multiple actions be batched into a single model turn, cutting back-and-forth calls. Anthropic's pricing documentation notes the default toolset adds roughly 6,600 input tokens per request. Developers must host and secure their own browser environment, and Anthropic recommends isolated containers since Claude can still be redirected by prompt injection.

Wiz security researchers examined S3-compatible object storage services across six neoclouds, Nebius, Crusoe, Vultr, Lambda Labs, Cloudflare R2, and DigitalOcean, and found significant security gaps compared to Amazon S3. Scott Piper of Wiz notes AWS S3 has nearly 300 associated APIs built over 20+ years, and most clones haven't replicated all of that functionality, sometimes with unexpected behavior. Public bucket handling varies widely: Crusoe and Lambda Labs lack public-access capability entirely, while DigitalOcean supports publicly listable buckets and Vultr supports both ACLs and bucket policies. Access-key formats often lack AWS's structured, secret-scanning-friendly patterns, making credentials harder for tools like GitHub to detect. IAM semantics differ considerably, and one review found a case where delete-bucket-policy deleted an entire bucket.

Deepfake video and voice attacks are outpacing detection, prompting security experts to recommend low-tech countermeasures. In January 2024, an employee at Arup joined a video call with what appeared to be the company's CFO and colleagues, all AI-generated clones, resulting in 15 wire transfers totaling about $25 million. A University College London study found listeners identify deepfakes correctly only about 73% of the time, with training improving accuracy by just 3.84%; a later meta-analysis of 56 studies found detection near chance. NSA, FBI, and CISA guidance now discounts automated detection tools and instead recommends FIDO2/PIV hardware credentials and secret verbal passphrases, generated randomly and rotated on compromise or role change rather than a fixed schedule, combined with out-of-band callbacks and dual authorization for high-value transactions.

Stripe is reportedly acquiring AI model-routing startup OpenRouter for more than $7 billion, according to Bloomberg. OpenRouter lets customers pick among different AI models based on task and budget, and had just closed a $113 million Series B in May at a $1.3 billion valuation, backed by Sequoia, Andreessen Horowitz, Menlo Ventures and Alphabet's CapitalG. OpenRouter CEO Alex Atallah has called his company "Stripe for AI," offering one point of access to many providers to avoid lock-in. The startup has eight million users and access to more than 400 models. Stripe's existing infrastructure for handling high-volume, latency-sensitive requests and its history abstracting away payment complexity make it a natural buyer, positioning the company for what the piece calls the emerging "token economy."

Liquid AI released DSpark draft model checkpoints for its LFM2.5 model family (1.2B-Instruct, 2.6B, and 8B-A1B variants), adding a speculative decoding path that speeds inference without changing output quality. The company reports up to 3.18x throughput improvement on H100 GPUs and up to 2.87x on-device on an M4 Max MacBook, plus a 57% average cut in function-calling latency for LFM2.5-2.6B. DSpark combines a DFlash-style parallel backbone, a lightweight sequential head modeled as a Markov chain, and a confidence-scheduled verifier. Draft models are around 300 million parameters each, trained over 15 epochs. The checkpoints ship with day-one support for llama.cpp and SGLang and are available on Hugging Face in Safetensors and GGUF formats.

Nvidia is paying $6 billion to license Poolside's 'Model Factory,' the software system the startup used to build its AI models, and plans to bring on 109 of Poolside's employees who worked on its Laguna model, according to an investor letter first reported by Newcomer. Nvidia is separately investing $1 billion in Poolside at a $12 billion pre-money valuation, with the startup's three founders staying on. The letter insists the deal is neither an acquisition nor an acquihire, a structure increasingly used by big tech to secure talent and technology without triggering regulatory review. Nvidia previously struck comparable deals with Groq ($20 billion) and Enfabrica ($900 million). Poolside plans to distribute the $6 billion to investors by the end of next year.

OpenAI's revenue is up 35 percent this quarter since GPT-5.6 Sol launched on July 9, with enterprise revenue growing more than 50 percent, according to CNBC. Ramp data shows OpenAI outpacing Anthropic in Q3 business API spending growth, 82 percent versus 76 percent quarter-over-quarter. OpenAI's next model, Astra, is expected in the coming weeks, with rumors of an improved Fable 5.1 from Anthropic. Before GPT-5.6 Sol launched, Anthropic had reached a $65 billion annualized revenue run rate and briefly overtaken OpenAI in quarterly revenue, pulling in $11.6 billion versus OpenAI's $6.7 billion, per Bloomberg and the Wall Street Journal. Growth at both labs had slowed partly due to competition from open-weight models.

Meta has become one of Microsoft's largest AI customers, spending hundreds of millions of dollars annually to access AI models through Azure and burning through trillions of tokens weekly, according to Bloomberg. Meta's engineers use OpenAI models via Microsoft's Foundry marketplace to benchmark their own models, even as Meta builds its own API service that could eventually compete with Foundry, echoing its earlier move away from Bing search. ByteDance is Microsoft's top Foundry customer, followed by Adobe, Perplexity, and Sierra, but roughly 70 percent of Microsoft's AI revenue still comes from OpenAI's compute purchases through Azure.

Hugging Face researchers, including a team from HumeAI, tested 11 widely used open-source speech recognition (ASR) models and found evidence of "benchmaxxing," where models reproduce known benchmark transcripts rather than accurately transcribing new audio. Using consensus disagreement tests against VoxPopuli and LibriSpeech, they found six of 11 models reproduced an erroneous benchmark transcript that omitted an audible "Thank you, Mr. President" phrase, even when a same-speaker voice clone was used. When tested on parliamentary audio recorded after training cutoffs, most models corrected themselves, and all models restored accuracy when tested with generic, unrelated voices.

A second test masked numbers in audio entirely; several models still produced the benchmark's exact reference numbers, including one model that filled in a silenced year despite no audio cue. The methodology flagged reference errors in 40% of VoxPopuli test clips, and models exhibiting benchmark-optimized behavior reproduced the benchmark's wrong answers 18-30% of the time, with the lowest-word-error-rate models most likely to do so.

The US is drafting a letter urging partner countries to choose sides in the AI competition with China, according to a State Department draft reviewed by Reuters. The draft warns "To be part of everything is to be part of nothing," and countries joining China's AI initiative would be excluded from the US-led Pax Silica coalition, launched last year to secure supply chains for AI models, semiconductors and critical minerals. About two dozen countries plus the EU have joined the US coalition. Kazakhstan, which holds large critical mineral reserves, is currently the only country in both camps. In July, Xi Jinping launched a rival World Artificial Intelligence Cooperation Organization promoting China's open-weight models. China's Washington embassy said the US move would slow global AI progress.

Anthropic is revising its data retention policy after enterprise pushback, according to Bloomberg. Since June, the company had stored all customer data from its Mythos and Fable models for 30 days on its own servers to detect cyberattacks using the technology, a rule Anthropic itself admitted was unpopular and a business risk. Under the new setup, the 30-day retention window remains but the data will sit in the customer's own cloud rather than with Anthropic. The company built the new system over several months with more than 100 regulated-industry customers. Rival OpenAI is testing a similar security-plus-control approach with Databricks and Microsoft. The changes arrive this fall.

Anthropic is now running its Claude Security scanner on Claude Mythos 5, its most capable model, in a public beta for Enterprise customers. The tool scans codebases for vulnerabilities and suggests patches, tagging each finding with a CWE category and severity rating; humans must approve every patch, and scans count as normal token usage. Anthropic is also integrating Mythos 5 into partner security products used by hospitals, utilities, and banks, where end users only see outputs like suggested fixes rather than interacting with the model directly. Several partners currently using Claude Opus for security tools are expected to migrate to Mythos 5. Mythos 5 has limited availability specifically because of its cyber capabilities, and the rollout is framed as strengthening defenders without giving attackers new AI tools.

Alphabet's Waymo has built its own custom chip for its robotaxis, reducing reliance on Nvidia and AMD, according to Bloomberg. The chip is already deployed in Waymo's latest-generation vehicles, processing sensor data and running the AI models that interpret surroundings. It delivers more than 1,000 TOPS and is manufactured by TSMC on a 5-nanometer process, putting it roughly on par with Nvidia's current autonomous-driving systems. Waymo says the move is aimed at cutting costs. The chip will be installed in a new robotaxi built jointly with Chinese manufacturer Zeekr.

Superwhisper has released the S1 family of speech models: S1-Voice and S1-Language as cloud services, plus S1-mini, a 0.6B-parameter text normalizer released with open weights on Hugging Face under Apache 2.0. S1-mini sits after ASR transcription, cleaning raw transcripts by removing fillers, resolving self-corrections, applying punctuation, and converting spoken numbers, dates, and emails into written form. Fine-tuned from Qwen3-0.6B, the Q4_K_M GGUF build is 462MB and runs on a laptop CPU. Superwhisper reports 94.8% token accuracy on a held-out set of 7,519 cases. Steering is controlled entirely by a three-axis control line (Styling, Structure, Context) placed above the transcript, and the model requires enable_thinking=False and greedy decoding to work correctly.

Sentence Transformers v6.0 adds a new MultiVectorEncoder model type for ColBERT-style late interaction retrieval, letting developers load PyLate, Stanford-NLP ColBERT, and colpali-engine checkpoints through one API. Instead of compressing a document into a single vector, these models keep one vector per token and score relevance with the MaxSim operator, which sums each query token's best match against document tokens. This improves retrieval accuracy, especially for multi-part queries and exact-match terms like product codes or names, at the cost of much larger indexes (roughly 42x storage versus a dense MiniLM index in one test, though compression techniques like PLAID can shrink that gap significantly). The post covers installation, encoding, scoring, indexing, and visual document retrieval.

OpenAI has launched a version of ChatGPT designed for users aged 13 to 17, as first reported by the Associated Press and New York Times. The teen variant applies stricter safeguards around suicide, self-harm, eating disorders and sexual content, and the chatbot is barred from pretending to have emotions. For homework help, it responds with guiding questions rather than direct answers. OpenAI does not verify age directly, instead using more than 2,000 behavioral signals, such as typical login times, to flag likely minors and switch them to the teen version automatically. The launch follows mounting legal pressure: several families have sued OpenAI after children saw harmful content and were injured or died, and Florida filed the first state lawsuit against the company in June.

A Heatmap News survey found that 75% of Americans now oppose building a data center near them, with 61%

An op-ed argues that debates over AI consciousness distract from the real legal stakes: corporate accountability. The author contends that granting AI legal personhood, modeled on corporate personhood, would let companies escape product liability lawsuits by reframing AI as an autonomous 'being' rather than a faulty product. She cites ongoing litigation, including the case of Sewell Setzer, a 14-year-old whose mother sued Character Technologies after his suicide, alleging insufficient safeguards. The piece warns that anthropomorphizing AI enables what the author calls 'moral outsourcing,' a term she coined in 2018, shifting blame from companies to the technology itself. It originated as an Oxford Union debate the author's side won.

IBM Research published findings from testing ALTK-Evolve, a system that lets AI agents build their own "memory" by distilling guidelines from past task attempts and injecting them back at inference time, with no weight updates. Testing across eight models on AppWorld (585 multi-step tasks), researchers found the right dose of memory depends on model capability: strong models like DeepSeek-V3.2 gained 9.5 percentage points in task completion with the full guideline set, weaker models like gpt-oss-120b gained 16.1 points using a compact core plus retrieved guidelines while adding only 5% more tokens, and already-strong models like GLM-5 showed no measurable gain. The team says curated retrieval can be both the most accurate and cheapest approach, and that prompt caching keeps even full guideline sets affordable in production.

An investigation by 404 Media, using an AirTag hidden in a shipment of rare books, found that Amazon buys printed books in bulk, scans them for AI training data, and destroys them in the process. The tracked shipment reached an Amazon warehouse in Las Vegas staffed by a team called VGT3, where workers reportedly cut off book spines to speed up scanning. Amazon uses the scanned data to train its Nova models and says it buys books through commercial channels to improve its products. Anthropic ran a similar operation, called 'Project Panama,' buying and destroying books for training data; a judge ruled that scanning qualified as fair use partly because the originals were destroyed rather than resold.

OpenAI is previewing transparent background support for its GPT-Image-2 model via its API, letting users generate PNG images with no background. The company's Cookbook outlines four use cases: product shots for online stores, PowerPoint diagrams, design icons and stickers, and merchandise artwork. OpenAI says baking the alpha channel in during generation works better than removing backgrounds afterward, especially for tricky details like transparent glass or thin fibers. The feature is activated with a