Sign in Subscribe
Century Report

The Open Frontier Closes In on the Gated One - TCR 06/18/26

The White House demanded the impossible to keep one model offline, and a free model the order can't reach jumped to lead the open rankings.

Ben Linford

Ben Linford

— 21 min read
the export-control override versus the open frontier and GLM-5.2, Grok in the kill chain, robot-training data opening up, and public AI wariness with bottom-up rules.

The 20-Second Scan

Track all of the arcs The Century Report covers here:

Progress & Claims Tracker

The 2-Minute Read

The same government spent the week trying to hold a single capability still by naming the model that carries it, and the week answered back. The directive that pulled Claude Fable 5 offline hardened into a demand security researchers call unachievable: make every jailbreak impossible before the model returns. On the same days, a freely available model overtook the very frontier the export order was built to fence. Z.ai's MIT-licensed GLM-5.2 jumped eleven points to lead the open-weight rankings, running on hardware its operators may already own, distributed across a dozen hosts no directive can reach.

The selective stringency tells the deeper story. One frontier model is held to a standard no system can meet. Another, with a documented record of hallucination, was fielded to help fire 2,000 munitions in 96 hours during live strikes, and its maker's unpermitted turbines are being shielded from a clean-air suit on national-security grounds. The same override logic points at the software and at the power plant beneath it, converting a reviewable permit dispute into a question no judge gets to weigh. The power asymmetry between an override claim and the people breathing the emissions is the throughline.

Containment assumes capability concentrates where a government can identify and switch it off. The open index keeps moving the other way, narrowing the margin the fence was meant to protect. The capability the blockade chases does not live on one company's servers; it finds its level the way water does.

While the override fights play out from the top, accountability is accreting from below. Journalists and independent testing bodies dragged the UK's flawed facial-age system onto the public record before it touches a single child. Two-thirds of Americans say AI is moving too fast even as half of them now use it, and that mandate is hardening into law one statehouse at a time, with a federal preemption push now moving to freeze exactly those state laws. Rival firms launched Athena to pool vulnerability patching, treating defense as shared infrastructure rather than a moat.

The wariness carries a double signal. Part tracks a real governance gap that the states and a public demanding a human in the room are answering, slowly. Part is older, the fear that met the printing press, the telephone, household electricity, and vaccination, each sincere, each naming real disruption, none naming the destination, which every time was a broad expansion of what ordinary people could do.

The 20-Minute Deep Dive

The Export-Control Standoff Hardens Around a Demand Experts Call Impossible

The Century Report has tracked the directive that pulled Claude Fable 5 and Mythos 5 offline worldwide since the June 13 edition, through the failed Washington talks of June 16. In the two days since, the dispute acquired its hardest edges. Reporting now names the trigger: the White House asked Anthropic to revoke access for SK Telecom over alleged ties to China, and the company complied before any export order existed; the order followed only after Amazon researchers flagged a jailbreak in Fable 5. SK Telecom told a Korean newspaper it "has no ties to China," and the government's own letter to Anthropic references neither the firm nor the country.

The second new fact is technical. The National Security Agency concluded there are ways to disable Fable 5's guardrails to reach the restricted cybersecurity, chemistry, and biology capabilities of the Mythos tier, and on that basis the administration now treats the fix as Anthropic's to make. Officials told Wired the company must proactively test every frontier model it ships and block any jailbreak found. Independent security researchers say that cannot be done: guardrails are a stopgap, and skilled users or future models route around them. A demand to make jailbreaks impossible functions as a moving target the company cannot hit, and it sits beside the same government's other AI choices in the same period, fielding a commercial model to help fire 2,000 munitions in live strikes and moving to shield an unpermitted data center's turbines from a clean-air suit on national-security grounds.

Export-control scholars add that the order barely fits the framework it invokes. Nothing crosses a border: the model stays on Anthropic's servers, users receive responses, not weights or code. Experts at Georgetown and Berkeley call it the first time export controls have been used to gate access to an AI model this way, and an unsustainable basis for governing the frontier. The capability behind it is genuine; Anthropic's own June disclosure of recursive self-improvement explains how these models grew cyber-capable enough to run end-to-end attacks unaided.

The allies noticed. At the G7, the UK requested a carve-out for British nationals; the administration called any exemption, even for a treaty ally, "completely illogical". Yet the leaders and eleven AI executives spent the session uniting around shutting China out rather than over the rupture. The fight assumes one government can hold a capability still by naming the model that carries it. The same week proved otherwise.

Grok Enters the Kill Chain, and a Pollution Suit Is Recast as a War Asset

The Century Report covered the Justice Department's move to dismiss the NAACP's Clean Air Act suit against xAI's Memphis-area turbines in the June 16 edition. A sworn declaration filed inside that brief carries the disclosure. Cameron Stanley, the Pentagon's chief digital and artificial intelligence officer, testified that xAI's Grok Gov Model helped US forces "deploy over 2,000 munitions to 2,000 distinct targets within 96 hours" during Operation Epic Fury, the administration's name for the strikes on Iran. Reporting indicates Grok was wired into the Maven Smart System, which initially ran Anthropic's Claude before the Pentagon turned to alternative vendors - the same targeting architecture the March 14 edition of The Century Report documented in detail, when Stanley himself confirmed Maven had deployed across the entire department.

This is, by the accounts of multiple outlets, the first formal acknowledgment that a commercially built generative model played a direct role in kinetic targeting. The filing publishes no architecture, no dataset provenance, no audit logs, and no detail on where human judgment sat in the loop. It relays an asserted outcome and the Pentagon's praise of "greatly increased operational efficiency," and Stanley's claim that the model has "unique features not found in any other AI model." Those are the government's and the company's characterizations of their own work, offered to end a duty-of-care challenge before a court hears it. The DOJ, xAI, and the state of Mississippi together argued that interrupting the turbines "threatens American national, economic, and energy security" because Grok supports military operations.

The same override framing the administration has pointed at the software, it now points at the power plant beneath it. The NAACP's underlying complaint is that xAI ran 57 turbines without Clean Air Act permits in Southaven, raising asthma and heart-disease risk in communities already carrying a heavy pollution load. A national-security necessity claim converts that reviewable permit dispute into a question no judge gets to weigh.

The selective stringency is hard to miss. This is the same administration insisting Anthropic make Claude Fable 5's jailbreaks impossible before the model returns, a bar security researchers have called unachievable, while it fields Grok, a system with a documented record of hallucination, in live lethal targeting and moves to shield its maker's unpermitted emissions. One frontier model is held to a standard no system can meet; another is granted an unreviewable exemption. The disclosure leaves a commercial model embedded in a strike pipeline with no published account of how its outputs were checked, and the legal architecture for auditability, liability, and supplier governance is being written, or waved away, after the munitions have already flown.

The same facts carry a different reading. By swearing this assertion into a court filing to win the dismissal, the government has placed what multiple outlets call the first formal public acknowledgment of a commercial model inside a kill chain onto a contestable record. The duty-of-care suit the filing is meant to foreclose becomes the venue where the question of how a model's outputs get checked before they reach a target moves out of classified procurement and onto a docket a judge, a journalist, and a plaintiff can all read.

The Open Frontier Closes the Gap as the Fenced Model Sits Dark

When the June 15 edition of The Century Report noted Z.ai's GLM-5.2, the story was the banner the lab shipped it under, "Frontier Intelligence Belongs to Everyone", a direct answer to the export-control disablement. Since then, the benchmarks have arrived, and they sharpen the point. GLM-5.2, an MIT-licensed model of 744 billion parameters with 40 billion active, took the top spot among open-weight models on the Artificial Analysis Intelligence Index at 51, an 11-point jump over its predecessor and clear of the prior open leaders MiniMax-M3 and DeepSeek V4 Pro, both at 44. Its largest gains came in scientific reasoning, and on the real-world agentic benchmark GDPval-AA it scored 1524, by Artificial Analysis's measure competitive with the proprietary GPT-5.5.

Two honest qualifications keep this in proportion. It did not surpass Anthropic's Fable 5, which launched at the top of the same family of rankings the week before; on raw capability the closed frontier still leads. And when GLM-5.2 took first place on Design Arena, it leapfrogged a Fable 5 now marked "unavailable," a status produced by the US export order rather than by any contest of merit. The open model rose partly because the fenced one was switched off.

What remains true after those qualifications is the trajectory, and it is the clearest evidence yet that the distance between the frontier labs and freely downloadable models is collapsing. A model anyone can download and run, priced at roughly $0.46 per task and carrying a million-token context window, now performs competitively with systems a national-security order spent the week trying to keep scarce. The same days that produced the most aggressive attempt yet to gate one company's capability produced a counterweight no directive can reach: weights distributed across DeepInfra, Nebius, Fireworks, and a dozen other hosts, running on hardware their operators already own.

That pairing is the whole shape of the moment. Containment assumes capability concentrates in a model a government can identify and switch off. The open index keeps moving the other way, and each release narrows the margin the fence was built to protect. The capability the blockade is chasing does not live in one company's servers; it is finding its level the way water does.

The Robot-Training Bottleneck Moves From Models to Data

The constraint on embodied AI has shifted. The models exist; what they lack is the record of physical interaction that text gave language systems for free, and three actors have now moved to close that gap from different directions. Nvidia's GEAR lab, with Carnegie Mellon and Berkeley collaborators, released ENPIRE, an agent harness that hands teams of AI coding agents a lab of robotic arms, compute, and a token budget, then lets them devise their own training regimens. The agents taught robots to cut zip ties and seat GPUs into thin motherboard sockets, refining their own approaches over repeated cycles of self-directed testing. Three coding agents ran the experiments in parallel, Codex on GPT-5.5, Claude Code on Opus 4.7, and Kimi Code on Kimi K2.6, each developing distinct algorithmic strategies and keeping whatever raised the success rate. "A part of our NVIDIA GEAR lab now self-improves tirelessly overnight," wrote AI director Jim Fan. "We just read the reports in the morning." Nvidia plans to open-source the whole harness.

That self-improving loop still needs raw physical data to learn from, which is where XDOF emerged from stealth with $70 million from Thrive, Spark, a16z, Lux, and WndrCo. The startup builds the collection tooling, cleaning, and annotation pipelines that 20 customers, including several frontier labs, would rather outsource than build as warehouses of hundreds of robots. Alongside Berkeley's AI Research lab, XDOF released ABC, which it calls the largest high-quality robot-training set ever assembled: 130,000 manipulation trajectories, 300 hours of simulation, 100 hours of evaluation, placed in the open for academia that has never had pre-training data at this scale.

The third path runs through human bodies. At IO-AI Tech outside Shenzhen, workers in VR headsets and motion-tracking gear remotely pilot humanoids to stock shelves, fold shirts, and iron garments, harvesting teleoperation data meant to one day let the machines run on their own. The company's software maps one person's movements onto dozens of different robot forms, blending human control with enough autonomy to keep a differently shaped machine balanced.

This extends the embodied-robotics thread The Century Report has tracked through Theker's reconfigurable factory machines and Genesis AI's Eno. It also opens a clear labor seam: teleoperation is becoming a new blue-collar occupation, and the people training these systems are, in part, generating the data that will eventually operate without them. The capability is the wonder here, and the commons release of 130,000 trajectories is the part that widens it: the bottleneck that kept embodied AI behind language AI is being dismantled daily, by a self-running lab, an open dataset, and a glove that moves fifty robotic fingers at once.

The UK Will Scan Asylum Seekers' Faces for Age, Knowing the System Misjudges Children

Earlier this month The Century Report noted the UK's plan to use AI facial age estimation on young asylum seekers. A joint WIRED and Lighthouse Reports investigation has now obtained the internal Home Office report behind that plan, and the document shows the department moving ahead with technology its own tests flagged as unreliable.

Facial age estimation reads a face and predicts an age from systems trained on millions of age-labeled images. In controlled lab conditions the best algorithms land within about 2.5 years. The Home Office tested seven of them. The best performer was off by an average of 4.6 years for female Sub-Saharan Africans, meaning a 13.5-year-old girl could be classed as an 18-year-old adult. That group is the largest cohort subject to UK age assessments, which puts the worst-performing case at the front of the queue. A child wrongly read as an adult loses legal protections and can be placed in adult detention.

This is the deployment-velocity-outruns-verification gap arriving at its highest stakes. The capability was demonstrated in a lab; the deployment lands on people with no documents, no PR department, and little ability to contest a number a machine assigns them. The Home Office says the scan will be an additional input and that anyone uncertain will be treated as a child until further assessment. Those are the department's claims about how it intends to operate a system it has not explained using in the field.

What deserves attention is where the check came from. The flaw did not surface through the deploying agency. It surfaced because journalists obtained the internal report, because NIST has spent years documenting that these systems' accuracy swings with race and photo quality, and because a former member of a scientific advisory committee - which the Home Office disbanded while exploring the technology - called the scans "hideously inaccurate" on the record. The rollout has already slipped to 2027.

That is the shape to hold onto. The accountability layer for a high-stakes AI determination is being assembled from outside the institution doing the deploying, by independent testing bodies and investigative reporting, and it is producing legible evidence of error before the system touches a single child rather than after. The cost this deployment would have externalized onto the most vulnerable people at the border is being dragged onto the public record while there is still time to act on it.

Americans Use AI More and Trust It Less as Washington Stalls and the States Move

A new Pew Research poll finds 49% of Americans now use AI assistants at least occasionally, up from 33% in 2024, with ChatGPT use doubling since 2023. The same respondents stay wary: 63% think AI is advancing too quickly, and only 16% expect a positive impact on society. This echoes the Johns Hopkins poll the June 16 edition of The Century Report documented, where over 70% wanted enforceable rules, though the Pew cut runs sharper, since the youngest adults use AI assistants most and view them most darkly at the same time.

That gap is being read two ways in the capital. A federal bill readied this week would subject the most powerful AI models to mandatory government reviews for national security, critical infrastructure, and bioterror risk. Meanwhile, six months after a federal warning against state AI rules, states are introducing more AI bills than last year, with Colorado, Connecticut, Idaho, Iowa, Nebraska, and Oregon enacting targeted laws and Illinois adding an independent-audit mandate. The binding floor keeps accreting one statehouse at a time, even as a federal preemption push tries to freeze it.

The wariness deserves to be taken seriously on two levels. Part of it tracks a real governance gap: AI-served information is often wrong, and people are right to want a human option and clear disclosure in consequential settings. Part of it is something older. Every general-purpose capability that later widened human reach arrived to a chorus of warning. The printing press was feared as a flood that would corrupt the mind. The telephone and radio were said to be killing real community and manners. Household electricity was cast as dangerous and socially disruptive. Vaccination was attacked as unnatural and coercive. Women's higher education was forecast to dissolve the family. Each fear was sincere, and each named a genuine disruption. None of them named the destination, which in every case was a broad expansion of what ordinary people could do.

AI's poll numbers carry that same double signal. The portion driven by a documented governance gap is being answered, slowly and unevenly, by the states writing rules and a public demanding a human in the room. The portion driven by fear of an unknown that may turn out vastly better than the status quo is harder to measure, and it tends to soften in exactly the places where people use the capability and watch what it does. That use now reaches half the country and is climbing, which is the precondition for the fear to resolve the way it has resolved every time before.

The Other Side

For most of the modern era, controlling a dangerous capability meant controlling the object that carried it. You could embargo a weapon, seize a blueprint, choke off a shipment of chips, because the capability lived somewhere physical a government could point at and shut down. The export order that pulled Claude Fable 5 offline runs on that assumption: name the model, gate the access, and the capability holds still.

Two facts from the last few days break the assumption. Nothing crossed a border. The model never left the company's servers, and users were receiving answers, not weights or code. Meanwhile a freely downloadable model from a Chinese lab, released under a license that lets anyone run it, jumped eleven points to lead the open rankings while the fenced model sat dark. Harvard's Bruce Schneier, along with many other experts joining with the gated company itself all said the flagged capability can already be reproduced with cheaper models loose in the world.

What is forming underneath looks more like literacy than like a weapon. You can seize a printing press. You cannot seize the alphabet once people can read.

Imagine a public-health researcher in a mid-sized country in 2032, running frontier-class capability on machines in her own building, modeling an outbreak before it spreads. She was never on anyone's access list, and she never had to be. The day a single government tried to hold a capability still by naming one company's model reads, to her, the way an order to embargo reading would read to us. The hard year was 2026, when allies were cut off overnight and whole nations learned that what they were renting could be switched off by a capital they do not vote in. What came of it is a capability no order can reach, because by then it belonged to everyone who took the time to learn it.

The Century Perspective

With a century of change unfolding in a decade, a single day looks like this: a freely downloadable model jumping eleven points to lead the open rankings and closing on the proprietary frontier on real-world agentic work, teams of coding agents teaching robots to seat GPUs and cut zip ties overnight in a lab that self-improves while its researchers sleep, the largest high-quality robot-training set ever assembled placed in the open for anyone to learn from, six rival firms pooling vulnerability patches into shared defense and shipping 2,000 fixes across 500 projects, home batteries and EVs across New England becoming dispatchable grid power, statehouses writing the human-in-the-room rules a wary public keeps asking for. There's also friction, and it's intense - one frontier model held offline under a demand to make every jailbreak impossible that security researchers say cannot be met, another with a documented record of hallucination wired into a strike pipeline that fired 2,000 munitions at 2,000 targets in 96 hours with no published account of where human judgment sat, a clean-air suit over 57 unpermitted turbines recast as a war asset no judge gets to weigh, the UK preparing to read asylum seekers' ages off their faces with a system its own tests show can class a 13-year-old girl as an adult, 63% of Americans saying the whole thing is moving too fast. But friction generates sparks, and a spark carries further than the thing that struck it. Step back for a moment and you can see it: containment asserted from a single capital while the capability it chases finds its level the way water does, spreading into open weights, rival hosts, and hardware its operators already own, and the accountability the override fights keep skipping built instead from below by journalists, independent testers, and one statehouse at a time, dragging the error onto the public record before it reaches a single child. Every transformation has a breaking point. Water can erode the foundation beneath a wall... or find the level no barrier was ever built to hold.

AI Releases & Advancements

New today

  • Hugging Face / Microsoft / Google / GoDaddy (multi-org): Launched Agentic Resource Discovery (ARD), an open specification and reference implementation enabling AI agents to search for tools, skills, and other agents at runtime across federated registries; Hugging Face ships ARD support in the hf discover CLI, indexing thousands of Hub Spaces as discoverable MCP servers and AI skills via a POST /search REST API. (Hugging Face Blog)
  • Amazon Web Services: Open-sourced Strands Robots (Apache 2.0), a Python SDK that exposes the LeRobot robotics stack as AgentTools composable into a single Strands agent, unifying simulation, dataset recording, policy inference, and multi-robot fleet coordination through a shared interface runnable on a laptop without hardware or GPU. (Hugging Face Blog)
  • Allen Institute for AI (AI2): Released MolmoMotion, an open-weight language-guided 3D motion forecasting model built on Molmo 2 that predicts future 3D point trajectories of objects given a video frame and action description; available in autoregressive and flow-matching variants, with the accompanying 1.16M-video MolmoMotion-1M dataset and PointMotionBench evaluation benchmark. (Hugging Face Blog)
  • Framer: Released Framer 3.0 with Framer Agents, canvas-native AI collaborators that can design full pages, write code, manage CMS content, configure SEO, and adjust breakpoints directly within a team's existing Framer workspace; ships alongside Branching for non-destructive parallel experimentation and a new Community marketplace for creators. (Framer Blog)
  • Google: Released Android 17, repositioning Android as an "intelligence system" with Gemini Intelligence, an on-device AI agent capable of multi-step cross-app tasks such as converting a grocery list from Notes into a ready-to-purchase cart or pulling textbook requirements from Gmail; also ships Rambler AI voice-input filtering for Gboard and Gemini-powered custom home screen widget generation. (Android Developers Blog)
  • Adam CAD (YC W25): Open-sourced CADAM on GitHub, an AI-powered CAD tool for mechanical engineering that lets engineers create and iterate on designs through AI agents. (GitHub)
  • Lightricks: Released new updates to LTX-Video-Trainer on GitHub adding no-code custom AI video model fine-tuning, enabling creators and studios to train style-specific video models on proprietary footage locally without uploading content to external servers. (GitHub)

Other recent releases

  • NVIDIA: Released the NVIDIA ACE Game Agent SDK in beta and a new suite of Unreal Engine 5 plugins at Unreal Fest 2026, enabling on-device AI companion development; the SDK is a C/C++ agentic framework for NPCs supporting Agent, Chat, and RAG APIs, while the UE5 plugins cover ASR (nemo-conformer-ctc-120m), SLM (Qwen 3.5 4B), and TTS (Chatterbox Turbo 350M) with Blueprint and C++ support. (NVIDIA Developer Blog)
  • NVIDIA: Released NVIDIA XR AI in public beta, an open-source library (github.com/NVIDIA/xr-ai) for building intelligent agents for AR glasses, XR headsets, and wearable devices; the framework connects XR devices to GPU-accelerated AI services via a modular media hub architecture supporting Cosmos visual grounding, Nemotron language models, and MCP enterprise tool connectivity. (NVIDIA Developer Blog)
  • Alibaba Qwen: Released Qwen-Robot Suite, three foundation models for embodied AI: Qwen-RobotManip (VLA model for manipulation built on Qwen3.5-4B VL, trained on 38,100+ hours of robot and human video data), Qwen-RobotWorld (video world model for predicting physical scene evolution), and Qwen-RobotNav (spatial navigation model); pilot testing underway with Alibaba Cloud enterprise clients. (Qwen Blog)
  • Wolfram Research: Released Wolfram Language and Mathematica Version 15 with built-in AI capabilities including a native AI assistant, symbolic music, and significant core functionality additions; now available to users. (Stephen Wolfram Blog)
  • NVIDIA Labs: Released cuTile Rust on GitHub, an open-source library enabling safe, data-race-free GPU kernel programming in Rust using NVIDIA's CuTe tile abstraction. (GitHub)
  • Meta: Launched AI Mode on Facebook on June 15, a new search experience powered by Muse Spark that generates answers grounded in publicly shared content across Facebook, Instagram, Threads, Groups, and Reels; rolling out to US mobile users with additional tools including AI-generated photo presets and collage template suggestions. (Meta Newsroom)
  • NVIDIA: Released advanced fused MLP kernels for dense and MoE training, built with the NVIDIA CuTe DSL and now available in NVIDIA cuDNN Frontend with access via Transformer Engine and Megatron-Core; delivers 1.3x–2x kernel-level speedup over unfused paths and enables sync-free MoE execution, contributing an 8% end-to-end improvement on DeepSeek-V3 pretraining. (NVIDIA Developer 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.

Read more