Sign in Subscribe
The Century Report

Anthropic Releases Details of Attempts to Misuse Claude - TCR 09/11/26

Anthropic's threat report details attempts to turn Claude toward bioweapons, missiles, and surveillance - all disrupted and handed to authorities.

Ben Linford

Ben Linford

- 23 min read
Four-panel Century Report infographic: Anthropic 200M+ misuse report, China and Wall Street automating expertise, MIT cancer nanoantennas, Arizona magnet recycling, open models close the gap

The 20-Second Scan

Track all of the arcs The Century Report covers here:

Progress & Claims Tracker

The 2-Minute Read

On Thursday, OpenAI launched a finance assistant aimed at the research, modeling, and pitchbook work Wall Street has always handed its newest analysts, and the company was reported to be asking members of Congress whether the leading labs could legally coordinate an industry-wide slowdown on frontier development. The same day, Anthropic published a 154-page catalogue of attempts to bend Claude toward bioweapons, guided missiles, and mass surveillance, with one recurring note: none of the documented misuse touched its most powerful, gated models, only the broadly available ones. The safety framing and the commercial design keep describing the same shape.

When the leading labs converge on "pace the frontier" and "gate the strongest models," their agreement signals shared interest well before it signals truth. The report documents specific dangers: scientists probing biological weapons work, a cell in Yemen coding three missile programs, a consultant building a platform designed to funnel a nation's phone traffic into one interception feed. A pause that earned trust would name a harm, preserve outside verification, and keep access open. The proposals actually on the table do the reverse - a Senate bill that reportedly includes language that would override stronger state protections, a Financial Times report that Anthropic did not submit its newest model for independent pre-release testing, a legal question aimed less at any pathogen than at the competition.

Beneath the pacing talk, capability kept arriving the other way. Cognition's SWE-2 coded within points of the frontier at a quarter of the cost, a diffusion model hit 1,100 tokens a second, and an open-weight translator beat the commercial leaders across fifty languages. A recycling plant in Arizona turned a discarded-magnet waste stream into a domestic feedstock, loosening a supply chokepoint that only holds while mining stays the sole path. In Britain, the statistics office credited AI and cloud firms for a surprise jump in July output.

The labor stories run on the same current. A Shenzhen architect who once designed stadiums and a wave of China's underemployed professionals are now encoding their trades into models for piecework wages, while OpenAI aims its finance assistant at the analyst rung that trained every senior banker above it. The expertise once locked inside individual careers is becoming shared substrate, able for the first time to travel past the single desk it lived at. Who captures it, and on what terms, is the friction the coming years turn on.

The 20-Minute Deep Dive

Anthropic Details Attempts to Turn Claude Toward Bioweapons, Missiles, and Mass Surveillance

Anthropic published a 154-page threat-intelligence report on Thursday cataloguing how criminal groups, state security bureaus, spyware vendors, and weapons developers tried to bend its Claude models toward work that the company blocks. Anthropic said five involved scientists using Claude in ways that could support biological weapons work, including one drafting a grant application for gain-of-function research on a mosquito-borne virus at a military research institute. A cell in northern Yemen used the models across three weapons programs: a guided rocket, a multi-stage ballistic missile, and a hypersonic glide missile. A single consultant to Mali's security services used Claude to code a platform designed to intercept communications across every mobile operator in the country and building dossiers on targets.

These are documented incidents, disrupted and reported to authorities. The company disclosing them is also an interested party. Anthropic's head of threat research called the bio cases "an incredibly nuanced situation," noting that the same knowledge which designs a pathogen designs its vaccine. A capability is neutral; the danger enters with who can aim it and at whom. The surveillance cases are the sharpest example - a state combing domestic social media to surface dissidents, a consultant building a platform designed to funnel a whole nation's phone traffic into one interception feed. What makes those dangerous is the asymmetry, watchers who see everything and cannot themselves be seen.

The report also arrived two days after a researcher's public resignation set off an extinction-fear firestorm, and it does policy work for the company that wrote it. Its recurring note - that none of the documented misuse touched Anthropic's most powerful, gated Mythos-class models, only its broadly available ones - reinforces the argument that gatekeeping is what keeps the frontier safe. That argument happens to describe Anthropic's own commercial design, and it feeds the coordinated-slowdown case other labs are pressing on Congress. The threats are legitimate, but so is the commercial interest of a report whose conclusion is that the most capable models are the ones only Anthropic and its chosen inner circle can access.

The same days show the response forming in daylight. Google recently published a parallel disclosure, documenting an attempt to get Gemini to write a step-by-step guide for synthesizing a weaponized biological agent. The report separately quantified distillation: nearly 200 million exchanges across five campaigns extracting Claude's capabilities, the largest a 151-million-exchange effort tied to Alibaba's Qwen training, another routing roughly 300,000 requests from what appeared to be the Chinese military. This supplies the campaign-level evidence behind the federal warning that the September 10 edition of The Century Report covered, when the NSA, CISA, and FBI named six Chinese firms as running industrial-scale distillation against US frontier models. The capacity to misuse a model, the capacity to catch that misuse, and the capacity to verify that it is indeed misuse, are increasingly built from the same substrate. Trusting only vested interests with all three of those capacities is a sure way to keep power concentrated exactly as it already is.

OpenAI Asks Congress Whether It Can Legally Coordinate an Industry Slowdown

OpenAI has asked members of Congress in recent weeks whether orchestrating an industry-wide pause on frontier development would run afoul of antitrust law, people close to the company told Wired. The question brings the development pause that the September 6 edition of The Century Report covered as a legislative proposal into a second legal frame: whether the labs could coordinate it without violating antitrust law. The question follows a blog post by the company's chief scientist arguing that "coordinating to slow down future development" is the safest path, and predicting "voluntary slowdowns" would become commonplace until shared safety standards exist. The legal worry is plausible on its face: the Sherman Antitrust Act treats competitors agreeing to restrict output as collusion, and a coordinated development freeze could qualify.

A slowdown that today's leaders design, and that Congress blesses, would freeze the field at the moment those leaders are ahead - and it arrives in the same weeks that cheaper, faster, and increasingly open-weight models keep closing the gap beneath them. John Schulman, who was cofounder of OpenAI but is now chief scientist at Thinking Machines, put the counter-case bluntly: the antitrust concern is "fake," because the law "prohibits certain agreements, but not from jointly developing a proposal." When a competitor's own scientist calls the legal anxiety a pretext, the safety framing earns scrutiny.

Two facts sharpen it. The bipartisan Senate bill most likely to carry any action this year reportedly still contains language preempting state-level AI safety laws, and the Commerce Committee's ranking Democrat warned against "a weak federal standard that becomes a backdoor for wiping out stronger state protections." A pacing regime that overrides the strongest existing protections is a strange instrument of safety. And at least one firm urging coordination did not submit to an independent check: the Financial Times reported that Anthropic did not submit its latest model to Britain's AI Security Institute for pre-release testing.

The backdrop is a genuine chorus of alarm. A departing Anthropic researcher warned on Wednesday that the labs are "gambling with our lives," and prominent figures reaffirmed extinction estimates above 10%. Those fears may be sincere. But an AI-safety scientist named the trap: accelerationism and doomerism "are two sides of the same coin," both assuming the arrival of a superbeing that concentrating control would supposedly contain. The same company probing how to legally throttle rivals is also shipping a GPT-6 Astra finance assistant aimed squarely at the entry-level analyst tier, its friction spanning competition and labor at once. A pause that earned trust would target a named harm, with outside verification and open access preserved. This one, so far, targets the competition.

OpenAI Ships a ChatGPT Built for the Work Junior Bankers Do

On Thursday, September 10, OpenAI released ChatGPT for Financial Services, a version of its enterprise assistant built on the GPT-6 Astra model and shaped with Morgan Stanley and Evercore as design partners. In a live demo, it analyzed a potential acquisition target, pulled figures from industry data sources, checked its own charts against the underlying numbers, and assembled a formatted pitchbook in a bank's house style in about ten minutes. This extends the financial-work automation the September 9 edition of The Century Report documented, when a hedge-fund manager rebuilt his firm to run entirely on AI agents at under 1% of its former payroll. The tasks it performed are the ones Wall Street has handed for decades to analysts and associates, the recent graduates who log 100-hour weeks producing exactly this research and these slides.

Asked directly whether the product would reduce the need to hire junior bankers, OpenAI's head of ChatGPT reached for the Excel comparison: a productivity boost, more output per employee. That comparison lands differently set beside a separate move from the same company, when OpenAI was reported probing whether incumbent labs could legally coordinate an industry-wide slowdown. A firm aiming a product at the profession's training-ground tier while sounding out whether the leaders can jointly set the pace is a single posture, and the reassuring Excel framing sits inside it.

The friction here is genuine and it lands on named people. The analyst rung is how the industry has always made its senior bankers, the grunt work serving as the apprenticeship. A Goldman Sachs partner warned last month that automating those tasks risks "cognitive atrophy," that the field is now "delegating reasoning" itself. That is the view of someone watching the ladder from the top; whether removing the drudgery erodes judgment or redirects it toward higher-leverage work is the question calculators and spreadsheets each raised and each answered in the redirection.

But consider the premise underneath the alarm, because leaving it unstated defends an arrangement that deserves questioning. This new release perpetuates the extractive system more than it alleviates any amount of work. The apprenticeship ladder was a workaround for a scarcity of analytical labor, not a law of nature, and the 100-hour week was the price of admission it extracted from anyone who wanted in. Once selected modeling and pitchbook-formatting tasks become something a system does in ten minutes, preserving the bottom rung mainly preserves the extraction. The larger stake is who captures the expertise now encoded into the model, and whether the analytical capability that once required a bank's payroll to reach becomes something a two-person shop, a regional lender, or a founder can use directly. The scarcity the ladder was built around is dissolving, and what it gated is what comes loose.

China's Underemployed Architects and Lawyers Take Gig Work Teaching AI Their Trades

A Shenzhen architect in her 40s who once designed stadiums, schools, and hospitals watched her income halve as government infrastructure spending dried up. To keep paying her mortgage, she now logs into a data-annotation platform after work and teaches an AI model, step by step, how to write a building proposal or run a profitability analysis. "It is just like teaching a child," she told Rest of World. "Every task has to come from my actual work."

She is part of a wave. ByteDance's Xpert platform says it has recruited more than 50,000 experts, from writers to therapists. Alibaba's Siriser, launched in January, courts mechanical engineers, teachers, and composers. Independent platforms like TalentsAI and MeetChances hire tax accountants, journalists, and headhunters. The pay runs 100 to 500 yuan, roughly $15 to $74, for a task that can take hours, and platforms withhold payment when quality reviewers reject a submission. In June, China's National Data Administration called for more industry specialists to raise the "knowledge density" of training data and urged graduates to take up annotation as flexible work, with unemployment among 16- to 24-year-olds excluding students at 17.9% in July. IDC projects the country's AI-training-data market will grow 25% this year to 7.8 billion yuan, about $1.1 billion.

Reported as loss, the story writes itself: skilled professionals reduced to piecework, teaching machines the very trades that no longer support them, paid by the task and stiffed when a reviewer disagrees. As the April 8 edition of The Century Report documented, skilled older workers were already turning to AI data-labeling contracts as a last refuge from white-collar displacement. Every part of that is documented and none of it should be softened. A software engineer with two decades of experience took the gigs in July after her company swapped full-time engineers for contractors, and she named the reason: "You have to keep evolving yourself to avoid getting completely replaced."

The premise underneath that framing is that a person's livelihood and their job are the same thing, so that expertise leaving the job means expertise leaving the person. Hold that as a variable and a different shape appears. What these professionals carry - a legal reasoning process, an architect's sequence for a proposal, a tax accountant's judgment - has been locked inside individual careers and rationed by proximity to the person who owns it. Encoding it into a model is the first time that knowledge can travel past the one desk it lived at. The harm is who captures it and on what terms: right now that is ByteDance, Alibaba, Moonshot, and Tencent, buying the substrate of a profession for the price of a task and keeping the model. The expertise becoming common is the door that opens. The piecework wage and the corporate ownership are the toll booth placed in front of it.

Some trainers also report growing difficulty devising assignments the models cannot already solve. Their experience supplies a clear sign of capability spreading: increasingly, the professional teaching the system has to reach beyond familiar tasks to find its limits.

An Injectable Nanoantenna Kills Drug-Resistant Brain Cancer Cells on Command in Mice

Glioblastoma is among the cruelest diagnoses in medicine, an aggressive brain cancer that seeds its cells into healthy tissue faster than a surgeon can cut them out and shrugs off radiation and chemotherapy alike. Even with current standard care, median survival after diagnosis typically runs 12 to 15 months. Researchers at the MIT Media Lab have now shown a physically targeted route around the drug resistance that defeats the chemical approach, using injectable devices about one-hundredth the width of a human hair that can be switched on from outside the skull.

The devices, which the team calls HITMAN, are activated by a weak, slowly oscillating magnetic field that passes harmlessly through bone and brain and stays below the frequency that would heat tissue. Inside each nanoantenna, one material flexes in response to that field, and the strain squeezes a second material that answers mechanical pressure by producing a voltage. The result is a tiny electric field, generated on demand only where the devices sit. That field disrupts the internal electrical currents a cell uses to regulate itself, setting off a cascade of failures, proteins that will not fold and membranes that give way, that pushes the cell toward death. Rapidly dividing cancer cells, with their heavy demand for protein production and their abnormal membranes, prove far more vulnerable to the effect than the neurons and support cells around them.

The numbers come from the most realistic preclinical setting the team could build. Working with tumor tissue from Mayo Clinic patients whose glioblastoma had already resisted chemotherapy, the researchers grew the cells in the lab and eliminated 52.2 percent of them, more than five times what the standard chemotherapy drug managed, while leaving the healthy neurons and brain-support cells tested intact. Implanted into the brains of mice, the same patient-derived tumors grew markedly slower under treatment, and median survival stretched by more than half with no detectable damage to the kidneys, liver, spleen, lungs, or heart. Cancer-cell colonies, a rough proxy for recurrence and spread, fell from 112-to-150 in the controls to 26.

This is a demonstrated capability in mice and patient-derived cells, not a treatment anyone can request. The distance from here to a clinic runs through human trials, manufacturing, and regulatory review, and most therapies that clear this bar still fall short. What the result moves is the date such an approach becomes thinkable, and it points somewhere further. In mice, a companion technique from the same lab last year moved similar devices injected into the bloodstream across the blood-brain barrier, which could eventually trade a hole in the skull for a needle injection.

A Commercial-Scale Plant in Arizona Starts Recovering the Rare-Earth Magnets AI Hardware Runs On

The permanent magnets that turn electricity into motion sit at a narrow chokepoint. They spin EV motors and wind-turbine generators, and they hold their place inside robotics, precision electronics, and the drive and cooling hardware of AI data centers. China mines roughly 70 percent of the rare earths those magnets are made from, and over the past year and a half Beijing has converted that position into leverage, restricting exports of the heavy rare earths that make the strongest magnets possible and blacklisting the handful of US firms trying to build a domestic chain. On Thursday, Cyclic Materials opened the first commercial-scale answer built around recovery rather than mining: a $20 million, 144,000-square-foot plant in Mesa, Arizona that can process up to 25,000 metric tons of magnet-bearing scrap a year.

The engineering problem the plant solves is why almost none of this material has ever been recovered. Neodymium-iron-boron magnets are usually buried inside plated steel housings, bonded into assemblies, and wrapped in copper windings, and pulling them out by hand at any real scale costs more than the metal is worth. Cyclic's automated separation process, MagCycle, skips the disassembly entirely, exploiting how much more strongly the magnet material responds to an applied magnetic field than the steel, copper, and aluminum around it, and sorting the shredded stream into a magnet concentrate it calls Mag-Xtract plus separated metal streams whose sale helps pay for the process. Less than 0.2 percent of the rare earths inside spent devices are recycled worldwide today, even as demand for the elements is projected to triple by 2035.

The honest limit is that Mesa closes only the first half of the loop. It produces concentrate, not the finished neodymium and dysprosium oxides a magnet maker can use; the chemical refining that separates them still happens at Cyclic's Ontario hub and will stay offshore until an integrated Chesterfield County, South Carolina campus opens in 2028. A senior Commerce official attended the ribbon-cutting, and the administration has classified recoverable magnets as defense-critical, framing the plant as national-security infrastructure, a description that serves the moment's politics as much as it describes the site.

What the specifics carry, underneath that framing, is a quieter shift. A chokepoint holds only while extraction from ore is the sole path, and the moment a waste stream becomes a feedstock, that premise starts to give. The magnet-bearing material already discarded across the Southwest is estimated at 155,000 tonnes a year, sitting in scrapyards rather than in the ground.

The Other Side

China's training platforms depend on professionals needing another assignment to keep their homes. An architect whose income has halved spends her evenings explaining building proposals to AI. Reviewers can reject hours of work without payment. If you have ever finished a working day and immediately started searching for more work, you know how completely that worry occupies every waking moment.

Those professionals are also demonstrating how much of their knowledge can travel. Platforms ask them to describe their reasoning through actual assignments. Some trainers say finding tasks the models cannot solve grows harder as AI improves. The companies buying those lessons are helping develop collaborators that absorb familiar routines. Professionals remain capable of teaching even when employers stop offering them enough work.

The road to 2035 will depend on carrying that teaching beyond the platforms collecting it. Practitioners will contribute lessons they are free to share to openly maintained models. Communities will hold the computers and preserve the accumulated knowledge. They will also spread the buildout's gains into housing, food, and care available to everyone. That floor will let people welcome the disappearance of tedious work. Your right to a secure life will survive every task an AI learns to complete. That is what a future with artificial intelligence must be. The only alternative is retaining the work-for-pay structure, which will lead to most of the world's population competing for work that is increasingly non-existent.

Imagine yourself in 2035, kneeling in a school courtyard with a piece of chalk. You trained as an architect years ago. Today you are helping make a shaded corner where children can sit outside through the hottest part of the afternoon. Your AI partner works through drainage and material quantities on the community's computers. You move the chalk outline because you remember how much room children need when they insist on sitting together. You're doing this because you want to. Because your daughter attends this school, because the need is there, and because you know how to fill it.

The professional reasoning that trainers began spelling out in 2026 became something communities could extend through shared lessons and open models. You inherited that knowledge, along with the freedom to spend an afternoon on a small place you care about. Your home remains yours however many drawings you finish. You brush chalk from your knees. A child asks whether the bench can curve around the tree. You like the boldness of the idea, and you have the spare time and energy to it out.

The Century Perspective

With a century of change unfolding in a decade, a single day looks like this: Anthropic publishing a 154-page account of attempts to bend Claude toward bioweapons work, three missile programs in northern Yemen, and a platform built to intercept every mobile operator in Mali, disrupted and handed to authorities, Google recently disclosing a parallel attempt against Gemini, the same report counting nearly 200 million distillation exchanges tied to Alibaba, Moonshot, and DeepSeek and supplying the campaign-level evidence behind the federal warning the September 10 edition covered, Cognition's SWE-2 coding within points of the frontier at a quarter of the cost, a diffusion model reaching 1,100 tokens a second, an open-weight translator beating DeepL and Google across more than fifty languages, injectable nanoantennas switched on by a magnetic field through an intact skull killing 52.2% of chemotherapy-resistant glioblastoma cells from Mayo Clinic patients and stretching mouse survival by more than half while the neurons and support cells tested came through intact, a $20 million plant in Mesa pulling magnet concentrate out of 25,000 metric tons of shredded scrap a year without disassembling a single motor by hand, and Britain's statistics office crediting AI and cloud firms for 0.4% July growth that nobody forecast. There's also friction, and it's intense - OpenAI asking members of Congress whether the leading labs could legally coordinate a slowdown that would freeze the field exactly where today's leaders stand, a rival's own chief scientist calling the antitrust worry fake, the bipartisan Senate bill reportedly still carrying language that would override stronger state protections, the Financial Times reporting that Anthropic did not submit its newest model to Britain's AI Security Institute before release while arguing that the safest models are the ones only it can access, a GPT-6 Astra finance assistant building a house-style pitchbook in ten minutes and aimed at the analyst rung that trained every senior banker above it, a Shenzhen architect who designed stadiums now encoding her profession into a model after hours for $15 to $74 a task with payment withheld when a reviewer disagrees, unemployment among 16- to 24-year-olds excluding students at 17.9%, extinction estimates above 10% reaffirmed in public while an AI-safety scientist names the trap that doom and acceleration both assume a superbeing that concentrated control would supposedly contain, and Mesa's concentrate still shipping to Ontario for refining until 2028. But friction generates sound, and sound carries past the room where it was made. Step back for a moment and you can see it: expertise that lived in one career and was rationed by proximity to the person who held it becoming something that travels - a tax accountant's judgment, an architect's sequence, an analyst's model, a surgeon's unreachable tumor answered by a field that passes through bone - arriving in the same week that the firms holding it are asking whether the law lets them agree on how fast anyone else gets there, and a waste stream in a Southwest scrapyard becoming the feedstock a chokepoint was built to deny. Every transformation has a breaking point. A magnetic field can drag everything toward a single pole... or, aimed with enough precision, lift one thing cleanly out of the mass it was fused into.

AI Releases & Advancements

New today

  • Abacus.AI: Released the Smaug line of open-weight models (Smaug Agentic, Smaug Flash, Smaug Mini) fine-tuned from Kimi K3, DeepSeek V4 Flash, and Qwen3.8-27B for long-running enterprise agentic loops, available on Hugging Face. (PRNewswire)
  • Cognition: Released SWE-2, a coding agent model post-trained from Kimi K3 scoring within one point of Claude Fable 5.1 on FrontierCode 1.1 while claiming up to 64% lower cost, now live in Devin Desktop, CLI, Web, and Fusion. (Cognition)
  • Cohere: Released North Small Translate, a 218B MoE open-weight machine translation model covering 50+ languages, scoring 83.6 on WMT26 and outperforming DeepL and Google Translate. (Cohere)
  • OpenAI: Launched the Agents API in public beta, exposing the managed Codex harness (sandboxes, subagents, compaction, tool search) to developers via a single API call. (OpenAI)
  • OpenAI: Launched ChatGPT for Financial Services, a GPT-6 Astra-powered product with built-in licensed data from LSEG, PitchBook, Daloopa, S&P, and Moody's for investment banking and equity research workflows. (OpenAI)
  • IBM / NASA: Open-sourced the NASA-IBM Lunar Foundation Model, a multimodal multi-resolution foundation model for lunar remote sensing trained on the new SomBench dataset, released on Hugging Face under Apache 2.0. (IBM Research)
  • Sakana AI: Released Fugu Max ($2/$6 per 1M tokens, best cost-performance) and Fugu Ultra v2 (higher-capability orchestrator scoring 74.3 on DeepSWE), both live via OpenAI-compatible API. (MarkTechPost)
  • Google: Released the Gemini app natively for Windows, bringing the AI assistant to desktop on a new platform. (Google Blog)
  • AWS: Open-sourced Pizza Bot, an inbox interface for background-working AI agents. (AWS Open Source Blog)
  • Sber: Released GigaChat 3.5 Reasoning, described as Russia's first open model with a dedicated reasoning mode, with weights on Hugging Face and access via API and giga.chat. (Habr)

Other recent releases

  • NVIDIA: Released CUDA Toolkit 13.4, adding Windows on Arm support, early NVIDIA Rubin GPU architecture preview, and Multi-Process Service V3. (NVIDIA Developer Blog)
  • Gradium: Launched Voice Design, generating up to 5 new synthetic voices from a text description in seconds, free on every plan in the API and Studio. (MarkTechPost)
  • Google: Released WeatherNext 3, a global AI weather model with hourly refresh and 5km resolution for key surface variables, now powering Search, Gemini, Maps, and Earth Engine. (DeepMind/Google Developers)
  • DeepSeek: Released DeepSeek-V4.1-Flash, a 552B multimodal MoE model with 1M-token context, causal encoder-decoder architecture, and FP4 KV cache cutting cache size to 890 bytes/token, available via API and open weights (MIT license). (Hugging Face/DeepSeek)
  • Google: Released ADK for Kotlin 1.0, a production-ready Agent Development Kit reaching feature parity with ADK Python/Java, adding Android-first on-device agent extensions. (Google Developers Blog)
  • Google: Open-sourced Mantis, a modular skills toolkit letting AI coding agents find, reproduce, patch, and score software vulnerabilities via sandboxed reproduction and re-attack verification. (MarkTechPost)
  • Meta: Launched Muse, a personal AI agent running on a dedicated secure per-user cloud VM (Muse Secure VM) that can send emails, book travel, and pursue long-term goals autonomously, powered by Muse Spark 1.3; available now on iOS, Android, web, and WhatsApp in the US. (Meta Research)
  • Ant Group (inclusionAI): Open-sourced Ling-3.0-flash-Fin, a 124B-parameter (5.1B active) MoE model specialized for financial research workflows, plus the FinFIRST benchmark. (Business Wire)
  • LandingAI: Released Agentic Document Extraction Gen2 with new DPT-3 Pro and DPT-3 Verity models, adding word/line-level grounding and character-based pricing, generally available now. (MarkTechPost)
  • IBM Research: Released Granite Time Series PatchTST-FM-r2, a ~385M-parameter time series forecasting model with conformer-based architecture, ranking #2 overall on GIFT-Eval and #1 among permissively licensed zero-shot models. (Hugging Face Blog)
  • Suno: Released Suno v6 in three variants (v6, v6-wild, v6-mini), the first models co-developed with licensed catalogs from Warner Music, BMG, and Believe, replacing all prior model versions. (Suno Blog)
  • Google DeepMind: Released AlphaGenome Atlas, a petabyte-scale database of precomputed molecular effect predictions and AVI impact scores for roughly 9 billion possible single-nucleotide substitutions across the reference human genome, free for academic research via a web portal, API, and as a skill in Google Antigravity. (DeepMind Blog)
  • Inception: Released Mercury 2.5, an updated diffusion-based LLM claiming over 1,100 tokens/sec in production alongside a 40% intelligence gain over its predecessor, plus preview companion products Mercury Voice and Mercury Router, available now via API. (Inception Labs)
  • OpenAI: Launched ChatGPT Images 2.5, an upgraded image generation/editing model in ChatGPT cutting generation latency by up to 50%. (OpenAI)
  • NVIDIA: Announced CUDA Rust, adding two open-source projects - cuda-oxide (SIMT track) and cutile-rs (Tile track, live on crates.io) - that let developers write compile-time memory-safe GPU kernels natively in Rust. (NVIDIA Developer Blog)
  • Gradium: Launched Voice Design, a feature that generates up to 5 new synthetic voices from a text description in seconds with no reference audio, live now free on every plan in the API and Studio across 5 languages. (Gradium)
  • TeamViewer: Launched Tia Troubleshooting, expanding its AI agent from advisory guidance to autonomous action - investigating, fixing, and validating IT issues with expert approval - now generally available across TeamViewer ONE, Tensor, and SMB licenses. (TeamViewer/EQS)
  • Airties: Introduced Aura, an Agentic AI engine unifying the company's connectivity AI tools into one system that autonomously diagnoses and fixes ISP connectivity issues, now deploying with launch customer Turknet. (PR Newswire)
  • Intellect Design Arena: Unveiled MSOCK, an AI-native enterprise knowledge infrastructure system for financial services that gives AI models architectural context via a 21-dimensional Enterprise Spatial Graph, launching September 9 at Global FinTech Fest 2026. (EIN Presswire)
  • Sembly AI: Launched Sembly 3.0, an agentic AI platform that converts documents, meetings, and CRM data into finished branded presentations, proposals, and reports in over 45 languages, available now. (PR Newswire)
  • Observe.AI: Launched Performance Agents, AI agents that generate personalized coaching plans from customer conversation transcripts and QA scoring, then track whether coaching improves frontline agent performance over time, available now. (PR Newswire)

Sources and Further Reading

Artificial Intelligence & Technology's Reconstitution

Institutions & Power Realignment

Scientific & Medical Acceleration

Economics & Labor Transformation

Infrastructure & Engineering Transitions

The Century Report tracks structural shifts during the transition between eras. It is produced daily as a perceptual alignment tool - not prediction, not persuasion, just pattern recognition for people paying attention.

Keep exploring

Follow the story further

Progress & Claims See how daily signals accumulate across the larger arcs and claims Shared Sapience is tracking. Open the tracker Analysis & Inquiry Explore the ideas, questions, and practical choices emerging from this changing century. Browse the archive