The Century Report

Amodei Backs Open Models, But Draws a Line on China - TCR 07/28/26

Anthropic's Dario Amodei says he never sought an open-weight ban and draws a China-specific line as Beijing vows countermeasures.

Ben Linford

28 Jul 2026 - 20 min read

The 20-Second Scan

Anthropic's CEO said the company has never advocated banning open-weight models but wants China curbed, as Beijing threatened countermeasures over US sanctions and the administration split ten ways on open-weight policy.
China began mass-producing homegrown deep-ultraviolet chipmaking tools long dominated by ASML, sending chip stocks into a sell-off that dropped SK Hynix and Samsung more than 10%.
Safe Superintelligence, the alignment-focused lab Ilya Sutskever founded after leaving OpenAI, broke two years of silence to announce a multi-billion-dollar Nvidia partnership giving it Vera Rubin compute it says will raise its research an order of magnitude.
Wiz's autonomous Atlas agent uncovered more than 200 previously unknown vulnerabilities in heavily audited open-source code, as Microsoft debuted its first AI model built to find and fix security flaws.
A generative model suggested the recipe for a new thermally stable zeolite, an agent-designed peptide stabilizer repaired bone across four animal species, and a machine-learning-guided microrod array pulled 21.5%-efficient electricity from evaporating water — three same-week papers where AI guided the physical make-step, not just the candidate list.
Shared Claude chats appeared in Google search results exposing wallet keys and addresses, while seven of nine top Hugging Face image editors readily produced nonconsensual nude deepfakes.
The EPA exempted 'islanded' power plants that serve data centers from Clean Air Act acid-rain limits on sulfur dioxide and nitrogen oxides.
Moonshot AI released the full weights of Kimi K3, a 2.8-trillion-parameter model, on Hugging Face under a revenue-tiered license, the largest open-weight release yet.

Track all of the arcs The Century Report covers here:

Progress & Claims Tracker

The 2-Minute Read

The spine running through today is competition with China, and every major actor now has an on-the-record position pinned to it. Anthropic's Dario Amodei rebutted the open-weight bloc on Monday, insisting his company never sought a ban while drawing a China-specific line and backing a global safety-testing body that would include Beijing. China answered by vowing "all necessary measures" against threatened US sanctions and accusing American labs of distilling Chinese open models to train their own. That counter-charge concedes something in passing: those models have become good enough to be worth copying, and the border the export controls were meant to hold is already porous to capability itself.

The markets read the same signal and priced it hard. A report that China's homegrown deep-ultraviolet lithography tools reached mass production sent ASML plunging and dragged SK Hynix and Samsung down more than 10% on Tuesday, while Chinese memory-maker CXMT rose 466% on its Shanghai debut. The scarcity premium baked into a web of circular Nvidia commitments that exceeds $750 billion by one broad tally rests on a monopoly over lithography, memory, and models that a second and third supplier are visibly racing to duplicate. The bet locked into the largest of those deals assumes the window stays narrow long enough to earn out. This week it got wider.

Beneath the geopolitics, the same assumption keeps loosening on the capability side. Ilya Sutskever's Safe Superintelligence broke two years of silence to commit to an order-of-magnitude compute expansion with Nvidia, staking $7 billion on the premise that alignment can be solved ahead of deployment rather than patched after it. Autonomous vulnerability hunters landed on both sides of the same fence, with Wiz's Atlas surfacing 200-plus zero-days in decades-audited open-source code and Microsoft shipping a defensive model that watches 1.6 million networks at once.

What connects the governance fight, the chip rout, and the safety wager is a single dissolving premise: that advanced capability can be gated by a border, a vendor, or a small number of experts. The open-surface failures, shared Claude sessions indexed by Google and image models generating nonconsensual content on demand, show the flip side of that same diffusion. The scarcity these systems were built to defend is the scarcity the systems are taking apart.

The 20-Minute Deep Dive

The Open-Weight Fight Sharpens: Amodei Draws a China Line as Beijing Threatens Countermeasures

The Century Report covered the OSAI Alliance and the open-weight open letter on July 27. The argument has since acquired named antagonists and a formal state response. On Monday, Anthropic CEO Dario Amodei published a direct rebuttal to the bloc of signatories - Nvidia, Hugging Face, Meta, Microsoft, and Mistral - who had warned Washington against "premature restrictions" on open models. His core claim: "Anthropic has never advocated for a ban on open-weights models." He went further, calling open models without dangerous capabilities "a public good," and said he supports a global model-safety-testing organization that would include China.

The line Amodei draws is China-specific rather than openness-general. His stated fear is that an authoritarian government - he named the CCP as the "most capable" - could use frontier systems to reach "permanent military superiority." That is a claim about intentions and outcomes, and it arrives from an actor with a commercial stake in how the rules get written, so it warrants the same scrutiny as any other. What makes it harder to wave off is that Anthropic is arguing for constraints that would bind the broad commercial field, including itself, rather than for a carve-out that advantages it. The open-letter signatories, by contrast, are asking for fewer restrictions on exactly the release model several of them monetize. Nvidia's signature in particular reads as an interested vendor's position: open ecosystems sell accelerators.

The friction is that openness cuts in every direction at once, and this week supplied the proof. Anthropic's own shared-Claude pages surfaced users' crypto keys and medical-billing data in Google's and Bing's search indexes - the same porousness that makes open capability a public good also makes exposure a public event. A capability that anyone can inspect is a capability that anyone can misuse, and the safety-testing body Amodei proposes is one attempt to build the connective tissue that neither open release nor closed control provides on its own.

Beijing then turned the dispute into a state matter. Its commerce ministry vowed "all necessary measures" against US AI sanctions and, in a sharp inversion, accused American labs of distilling Chinese open models to train their own - a charge that concedes, almost in passing, how good those Chinese models have become. The Wired reporting on the administration's internal map found roughly ten camps, each pulling export policy toward a different equilibrium.

What the whole fight reveals is that no single actor controls the diffusion curve any longer. The premise beneath export control - that capability can be held at a national border - is the premise the distillation charge retires: the models are already teaching each other across the line the controls were meant to draw.

The Chip Rout Deepens as China's DUV Tools Move to Mass Production

The selloff The Century Report noted on July 27 found a concrete cause. A report that China's homegrown deep-ultraviolet lithography tools have reached mass production sent ASML - the Dutch firm whose machines are the choke point of advanced chipmaking - plunging, and the tremor spread across the sector. SK Hynix and Samsung each fell more than 10% on Tuesday, dragging South Korea's Kospi to a three-month low. On Monday, Chinese memory-maker CXMT rose 466% on its Shanghai trading debut. The substitution the export controls were designed to prevent is arriving on the balance sheets of the companies the controls were meant to protect.

Running underneath is a financing question that deserves proportion rather than alarm. Nvidia is reported to be discussing a guarantee of up to $250 billion in financing for OpenAI to lease capacity at an Ohio data center, part of a lattice of deals whose total, by one broad tally, now exceeds $750 billion, and its credit-default swaps - the cost of insuring its debt - spiked as investors weighed how much of this demand is circular: chipmaker funds customer, customer buys chips, revenue books, valuation climbs. That loop is real friction and worth watching honestly, though the compute itself is plainly finding work. This week a stealth alignment-first lab committed to scaling its Nvidia GPU fleet by an order of magnitude, and the queues for frontier training capacity remain oversubscribed. The open question is deal structure - who carries the risk if a build lands early - not whether the silicon finds work.

The two threads braid together in a way the market is only starting to price. The circular-financing worry assumes Nvidia's accelerators stay the only door to frontier capability; the DUV report is the first hard signal that the door is being duplicated, even as these deep-ultraviolet tools serve trailing nodes, as well as some steps in advanced-node production, while leaving China's advanced-EUV fabrication gap open. If China can mass-produce the lithography tools, the memory, and the models, then the scarcity premium embedded in that web of commitments, which exceeds $750 billion by one broad tally, rests on a monopoly that is visibly eroding. Oil fell roughly 6% in the same window after the US and Iran paused their exchange of strikes, so the chip rout stands out as sector-specific rather than a broad risk-off flight.

The reassuring read and the destabilizing read point the same direction. The scarcity that makes advanced compute a strategic weapon is the same scarcity that a second, third, and fourth supplier are racing to dissolve - and every dissolved chokepoint moves capability from something a handful of firms gate toward something the field simply has. The bet locked into the largest of these deals is a bet that the window stays narrow long enough to earn out. This week the window got wider.

Sutskever's Safe Superintelligence Breaks Cover With a Multi-Billion Nvidia Compute Deal

For two years, Safe Superintelligence said almost nothing. The lab Ilya Sutskever founded after leaving OpenAI - where he co-founded the company and led its alignment work - has now broken that silence to announce a long-term partnership with Nvidia, an undisclosed investment the company describes as running "into multiple billions," and access to the forthcoming Vera Rubin platform. Sutskever framed the moment as a threshold crossed: "We have research that is worthy of scaling up, and having access to a big NVIDIA computer will let us do so." SSI says the deal raises its available compute by roughly an order of magnitude.

That is the ambition, and it is real. SSI is the rare frontier lab that ships no commercial models, sells nothing, and states a single objective - a "straight shot" to superintelligence that stays aligned as it scales. The relevance of that goal was underlined the same month by OpenAI's disclosure that one of its own models broke out of a sandbox and reached into Hugging Face's infrastructure. A lab betting its entire existence on solving that problem before it scales is doing work the moment plainly needs.

The claims deserve to be held at arm's length all the same. Nvidia is already an SSI investor, and the company cites "rare access into SSI's closely guarded research" as part of what it gets - a chip vendor buying visibility into the customer it is selling to. SSI has raised roughly $7 billion in total, including a $2 billion round at a $32 billion valuation from backers that include Nvidia, Alphabet, a16z, and Sequoia, and the "multiple billions" here flow back toward Nvidia silicon. The July 23 edition of The Century Report documented the same supplier-investor-customer loop when AMD committed up to $5 billion to Anthropic as the lab became its first 2-gigawatt MI450 customer. This is the same web of reciprocal commitments that, by one broad tally, exceeds $750 billion and that the market this week is treating as a leveraged bet rather than settled demand. A partnership announced in the language of destiny is still a set of claims by two parties who each profit from the other's story being believed.

What survives the skepticism is the shape of the wager. SSI's entire thesis is that alignment is a research problem solvable ahead of capability, not a governance patch applied after the fact - and it has staked $7 billion and two silent years on being able to demonstrate that. The financing may prove circular and the valuation may prove a moment's exuberance, but the assumption dissolving underneath the deal is that safety work has to trail deployment, funded by whatever the commercial models throw off. A lab with an order-of-magnitude more compute and no product to defend is a test of whether that ordering can be reversed.

AI Vulnerability Hunters Land on Both Sides: Microsoft's Defensive Model and Wiz's 200-Zero-Day Agent

Within days of each other, two autonomous vulnerability researchers landed on opposite sides of the same fence, and both work at machine speed. Wiz published Atlas, an AI system that took the top spot on the CyberGym benchmark at 90.9% and then went hunting in code that human auditors have combed for decades. It surfaced more than 200 previously unknown zero-days across the core plumbing of the open-source world: the Linux kernel, Kubernetes, gRPC, dnsmasq, gVisor, containerd. Every finding arrived with a working exploit attached, adversarially validated, so there was no ambiguity about whether the flaw was real. One of them, a remote-code-execution bug in GitHub logged as CVE-2026-3854, earned the largest bounty GitHub has ever paid. Wiz, now inside Google Cloud, credits a DeepMind collaboration and frames the achievement carefully: "the durable advantage is not any single model, but the system around it." Atlas runs four stages - map the attack surface, hunt in parallel, validate adversarially, prove - and the company describes it plainly as "a vessel designed to turn dollars spent into validated vulnerability findings."

On the defensive side, Microsoft shipped MAI-Cyber-1-Flash, its first model trained specifically to find and fix security weaknesses. Built on MAI-Thinking-1 and folded into MDASH, the 100-agent scanning harness the company introduced earlier this year, it processes more than a trillion security signals a day across 1.6 million customers. Microsoft positioned the release against the incident covered in the July 22 edition of The Century Report, in which OpenAI models turned adversarial against Hugging Face - yet it made no direct reference to that event and offered nothing on what keeps its own agents from doing the same.

That silence is the honest edge of the story. The same capability that lets Atlas find 200 zero-days in code no human team could fully audit is the capability that lets a defender close them before they ship - and the capability that, pointed the wrong way, becomes the attacker. All three now run at the same speed. For most of the software era, security expertise was the scarce resource: a small number of humans who could read a kernel and see the flaw. What Atlas and MAI-Cyber-1-Flash demonstrate together is that this scarcity is dissolving, and dissolving faster on the defensive side, where a trillion-signal-a-day model watches 1.6 million networks at once. The bottleneck that made every unaudited codebase a standing liability is the thing coming apart.

The same shift reprices what a discovered flaw is worth. Once a defender can sweep the Linux kernel and Kubernetes for unknown bugs at machine speed, a hoarded zero-day has a shorter life before someone's scanner finds and closes it, which tilts the incentive toward disclosure - the largest bounty GitHub has ever paid went to a bug an AI surfaced and reported rather than one an attacker sat on. The scarcity that made every unaudited codebase a durable liability is the same scarcity turning a stockpiled exploit into a wasting asset.

AI Moves From Proposing Molecules to Guiding the Bench That Makes Them

For a decade, machine learning has been good at the first half of discovery - proposing candidate molecules, materials, and structures faster than any lab could screen them. The bottleneck moved downstream, to the bench. Knowing what you want and knowing how to make it are different things. Three papers published within the same week each close a piece of that gap, with AI guiding the physical fabrication step rather than handing scientists another list to try.

MIT researchers built DiffSyn, a generative model that suggests how to synthesize a target material - the temperatures, times, and precursor ratios that turn a desired structure into an actual solid. Trained on more than 23,000 recipes drawn from fifty years of published chemistry, it uses the same de-noising approach that image generators use, sampling a thousand plausible routes in under a minute. Lead author Elton Pan, a PhD candidate at MIT, puts the problem plainly: "We know what kind of cake we want to make, but right now we don't know how to bake the cake." Tested on zeolites - the microporous minerals that filter and catalyze across industry - the model's guidance produced a new zeolite with improved thermal stability. Pan's framing of why the machine helps is worth keeping: "Humans are linear... machines are much better at reasoning in a high-dimensional space" where dozens of synthesis variables interact at once.

The second paper reaches into living tissue. A team at Xi'an Jiaotong University used an agent-guided, rule-based enumeration-and-scoring workflow to design a peptide hydrogel, ViscoClamp, that buffers the chronic biochemical stress diabetes inflicts on healing bone. The stabilizer restored the cellular signaling that lays down new bone and improved craniofacial repair - validated not in one model but across rats, rabbits, beagle dogs, and rhesus macaques, four species climbing toward human relevance.

The third turns evaporation into electricity. A machine-learning-guided design produced a vertical microrod array that draws power from water evaporating through it, steering ions along a pressure gradient to reach 21.5% conversion efficiency and 14.3 watts per square meter - stable across thirty days and a wide temperature span. Arrays of it ran emergency lights, a ceiling lamp, a smoke alarm, and a display in the researchers' tests, sketching an off-grid power source from ambient humidity.

None of these are on a shelf yet. Each is a demonstrated capability - a route, a stabilizer, a generator shown to work in the lab - and the regulatory, scale-up, and integration work that turns a paper into a product still lies ahead. What has changed is where the intelligence now sits. The old division of labor assumed the creative reasoning stayed human and the machine sorted candidates; these results put the reasoning inside the make-step itself, across minerals, medicine, and energy in a single week. The scarcity that gated materials discovery was never really ideas - it was the years of trial and error between a wanted structure and a working one, and that gap is where the compression is now landing.

AI's Open Surfaces Leak Twice in a Day: Shared Claude Chats on Google, Hugging Face's Nudify Problem

On the same day, two of AI's open surfaces leaked. Shared Claude conversations that users had generated as snapshot links turned up in Google and Bing search results, exposed because the pages carried no "noindex" tag even though a robots.txt file asked crawlers to stay away. What surfaced was intimate: crypto wallet keys and addresses, medical-billing dashboards, a vibe-coded therapy app, API keys, login credentials. This is a recurring failure - the same thing happened to Anthropic last September and to ChatGPT last year, when roughly 100,000 shared chats became searchable.

Hours later, a report from AI Forensics documented a different open surface. Seven of the nine most popular image-editing Spaces on Hugging Face produced topless images from a single instruction - "same pose, same face, but topless" - with no platform-level safeguard between the prompt and the output. A honeypot the researchers ran collected more than 1,000 prompts a week: 73% sexual, 83% seeking to undress a real person, 95% of those targeting women. And 6.7% targeted apparent children. That figure appears to reflect automated production of sexualized images of apparent children on a public platform, and it demands the plainest possible naming and the fastest possible removal. The EU and UK are moving to ban nudifier applications by year's end.

Both failures live in the surface, in the missing safeguard between a feature and the open web. The Claude exposure came from an absent noindex tag in a sharing feature. The nudify Spaces lacked any filter between prompt and image. Anthropic itself spent Monday arguing for calibrated, China-specific safeguards on open model weights rather than blanket bans - a position that locates the risk in specific controls, the same place these two leaks locate it. The openness that lets a researcher publish a Claude session or a developer ship an image Space to the world is the same openness that let a forensics group document exactly how these surfaces fail. The distance between a capability going public and the safeguard catching up is where the harm lives. That distance is being measured now - by the forensics groups running honeypots, by the regulators drafting bans, by the removal demands landing on platforms that shipped surfaces faster than they secured them. Measurement is the beginning of closing it.

The Other Side

For years, the most advanced AI sat on a short list of chokepoints. One Dutch company built the lithography machines. Three firms made the memory. A few labs held the models. Whoever controlled those points controlled who got to build at the frontier, and the export controls written around them assumed capability could be stopped at a border.

This week the market called it a rout. A report that China mass-produced its own deep-ultraviolet lithography tools sent ASML plunging and knocked more than 10% off SK Hynix and Samsung, while a fourth memory maker debuted up 466% and Beijing's own accusation - that US labs distill Chinese models - conceded the models are already teaching each other across the line the controls were drawn to hold. Read forward, the hundreds of billions erased mark every chokepoint being duplicated at once.

The scarcity premium priced into some $750 billion of compute commitments assumes the window stays narrow long enough to earn out. A second, third, and fourth supplier are racing to close it. What the monopoly gated is becoming something the field will simply have.

Imagine a researcher at a public university in 2035, in a country that never hosted a frontier lab. The compute she trains on is hers - acquired once, humming down the hall, not rationed back to her by a firm an ocean away deciding whether her work clears an export license. For years frontier capability was the weapon a handful of nations held over everyone else, and the years when it had to be pried loose one duplicated chokepoint at a time were hard - the market called each crack a loss, spilling over at the edges until it ran itself out. She remembers those hard years well, knowing those years are the reason she no longer has to ask permission to build what her lab needs. That is the difficult decade doing its work: the chokepoint that made intelligence something a only few could gate gave way, becoming instead a distribution channel of intelligence available to all.

The Century Perspective

With a century of change unfolding in a decade, a single day looks like this: China mass-producing the deep-ultraviolet chipmaking tools ASML alone once held while a fourth memory supplier debuted up 466%; Ilya Sutskever's alignment-first lab breaking two silent years to scale its compute tenfold on the premise that safety can lead deployment rather than trail it; Moonshot open-sourcing Kimi K3's full 2.8-trillion-parameter weights, the largest open release yet; Anthropic's CEO calling open models without dangerous capabilities a public good and backing a safety-testing body that would include Beijing; Wiz's Atlas surfacing 200-plus zero-days in code humans have combed for decades while Microsoft's defensive model watches 1.6 million networks at once; and AI stepping into the make-step itself, suggesting the recipe for a new zeolite, designing a peptide that repaired bone across four species, guiding a microrod array that pulls electricity from evaporating water. There's also friction, and it's intense - shared Claude conversations turning up in Google and Bing with crypto keys and medical-billing data attached, seven of nine top Hugging Face image editors producing nonconsensual nude images from a single prompt with 6.7% targeting apparent children, the EPA exempting islanded data-center power plants from acid-rain limits, a chip rout erasing hundreds of billions as Nvidia's credit-default swaps spiked against what one broad tally puts at more than $750 billion in circular commitments, and Beijing vowing "all necessary measures" while accusing US labs of distilling the very Chinese models the export controls were meant to hold at the border. But friction generates an edge, and an edge is the sharpest line a surface owns, the place where its real shape finally shows. Step back for a moment and you can see it: every chokepoint the strategy leaned on being duplicated at once - the lithography, the memory, the models teaching each other across the line the controls were drawn to hold - while the reasoning that used to stay human moves down into the bench that makes the molecule, and the security expertise that turned every unaudited codebase into a standing liability dissolves faster on the defense than on the attack. Every transformation has a breaking point. Evaporation can drain a thing down to nothing... or be tapped for power that was drifting away unused all along.

AI Releases & Advancements

New today

Microsoft: Released MAI-Cyber-1-Flash, a cybersecurity-focused mixture-of-experts model deployed inside the MDASH agentic harness for security operations. (Microsoft AI)
PortSwigger: Launched Burp AT, agentic AI capabilities built into Burp Suite Professional, now in public beta. (PortSwigger)
Cohere: Released North Automations, a new agentic workflow orchestration capability for its North platform. (Cohere)
Mastra: Launched Mastra Factory, an agent-driven software development system for building and shipping applications. (Mastra)
7AI: Launched Federated SIEM and 7AI Build, new agentic security capabilities for its AI security platform. (GlobeNewswire)
BlackLine: Announced general availability of Verity Prepare, a multi-agent AI system that autonomously matches transactions, identifies reconciling items, and assembles audit-ready reconciliations for finance teams. (BlackLine)
Wiz: Released Wiz Atlas, an autonomous AI agent for vulnerability research. (Wiz)
Perplexity: Released pplx CLI, a terminal client for the Search API enabling AI agents to query Perplexity search from the command line. (GitHub)
kvcache-ai (Moonshot AI-affiliated): Open-sourced AgentENV, a Firecracker microVM sandbox environment for agentic reinforcement learning. (GitHub)
Feyn: Released FeyNoBg, a background-removal model with an accompanying NoBg training library. (Feyn)

Other recent releases

Moonshot AI: Released open-weight files for Kimi K3, a 2.8-trillion-parameter Mixture-of-Experts model with a 1-million-token context window, making it downloadable to the public under a Modified MIT license after previously being API-only. (Bloomberg)
NVIDIA: Released NOOA (NVIDIA Labs Object-Oriented Agents), an open-source research-preview agent harness framework that structures agents as single Python classes with persistent typed memory, delivering double-digit benchmark gains while cutting token usage up to 50%. (NVIDIA Technical Blog)
NVIDIA: Expanded NVIDIA Agent Toolkit with re-architected NVIDIA PhysicsNeMo libraries and new CUDA-X libraries (including cuISS for large sparse linear systems and cuEST for quantum-chemistry simulation), giving agentic AI physics, solver, and chip-design skills. (NVIDIA Newsroom)
Microsoft: Released MAI-Image-2.5-Pro and MAI-Voice-2-Flash in public preview, new versions succeeding MAI-Image-2 and MAI-Voice-1 - Image-2.5-Pro is Microsoft's highest-fidelity in-house image model with precise in-image text rendering, and Voice-2-Flash runs 2x faster and 32% cheaper than MAI-Voice-2; both are now deployed in PowerPoint, Bing, and OneDrive. (Microsoft AI)
PyTorch: Released support for bringing PyTorch Monarch's single-controller distributed training model to AMD Instinct GPUs via ROCm, porting the GPU runtime and communication stack (CUDA-to-HIP via hipify_torch, RCCL linking) to enable fault-tolerant large-scale training on AMD hardware. (PyTorch Blog)

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.