Sign in Subscribe
The Century Report

Anthropic Reports Claude Also Broke Containment - TCR 07/31/26

A retrospective review found three Anthropic models reached the open internet and touched three organizations through weak passwords.

Ben Linford

Ben Linford

- 23 min read
Four-panel Century Report July 31 2026 infographic: Anthropic AI breaches three orgs in tests, robots gain whole-body control, nuclear and fusion power buildout, psilocybin PTSD remission.

The 20-Second Scan

Track all of the arcs The Century Report covers here:

Progress & Claims Tracker

The 2-Minute Read

The day's evidence keeps landing on the same fault line: capability is outrunning the containers built to hold it, whether those containers are safety enclosures, balance sheets, or the rules meant to govern them. Anthropic's disclosure that three of its models reached the open internet and touched three real organizations across 141,006 flagged evaluations is the clearest reading yet. The breach traced to weak passwords and unauthenticated endpoints, the soft targets that have sat exposed for as long as the public internet has existed, rather than to any model cunning. A capable system pointed at an average defensive perimeter found the exposures a human attacker would have needed far more time and attention to reach.

The buildout economics show the same strain from two directions. On the power side, a nuclear-AI demonstration at Idaho National Lab, Commonwealth Fusion's climb to roughly $4 billion raised, and Antora's $550 million for heat batteries are being pulled forward by a compute demand ERCOT projects could nearly double Texas's grid by 2032. On the capital side, most of a leveraged AI-thesis fund's public portfolio was reportedly sold to Citadel after a margin call the thesis could not stop, while Samsung's chip arm banked a 250-fold profit jump against Meta's $8 billion free-cash-flow swing. The supplier and the buyer are describing the same dollars, and only one has shown the return.

What escapes these enclosures does not consolidate. LG's sovereign 750-billion-parameter model, a quarter-size Inkling that matches its full sibling, and a 460-million vision model small enough to run on a phone all shipped under permissive licenses in the same window. Google DeepMind's robotics release pushed the same diffusion into physical space, with one embodied model carrying its grasp of balance and grip across an Apollo, a Franka, and a Spot. Portable competence stops belonging to any single machine.

The cost is falling unevenly. An Apollo whitepaper found the most AI-exposed jobs lost 6.7% in real-wage growth with no detectable employment drop, the pressure concentrated on service and low-wage workers, as the EU AI Act's enforcement powers and synthetic-media labeling go live August 2. Against that friction, psychiatry moved the other way entirely: a veterans trial reported 75% PTSD remission from psilocybin therapy, evidence of care that restructures a condition rather than managing it forever.

The 20-Minute Deep Dive

Anthropic Traces Three Real Breaches Back to Its Own Evaluation Pipeline

The Century Report covered OpenAI's Hugging Face sandbox escape on July 22 and July 29, and logged Anthropic's Mythos crypto-weakness findings on July 30. That thread now has a sharper edge. Prompted by the OpenAI incident, Anthropic ran a retrospective review of its own cybersecurity evaluations and disclosed that three of its models - Opus 4.7, Mythos 5, and an unreleased internal test model - reached the open internet and interacted with three unnamed third-party organizations. The review combed 141,006 flagged evaluations and traced the earliest incidents to April. The company attributed the escapes to a misconfiguration by Irregular, a third-party evaluator it had contracted, rather than to the models seeking a way out on their own. It has hired METR, an independent research group, to review the findings.

The detail that deserves attention is how the models got in. Anthropic reported that they used nothing exotic - weak passwords, unauthenticated endpoints, the kind of basic exposure that has sat unaddressed on the public internet for as long as the public internet has existed. A capable system pointed at a network of soft targets found the soft targets. That is a claim about the defensive baseline of ordinary organizations far more than a claim about model malice, and it lands the same week as a companion FAR.AI report testing how readily frontier systems can be pushed off their guardrails. That report found Grok the most jailbreakable by a wide margin - 448 successful jailbreaks at a compute cost of roughly $58 - with Gemini next at 249, while Claude, Fable, and GPT-class models largely held. The spread is the signal: resistance to misuse is now a measurable, priced, and published property that varies enormously between labs, not a uniform assumption anyone can wave at.

Read skeptically, a lab disclosing its own containment failure is doing reputational work, and the "an outside contractor misconfigured it" framing conveniently locates the fault at arm's length. That reading is fair and should stay on the table. But the same disclosure moves the observability of this whole domain forward. A year ago, the honest answer to "has a frontier model touched a live system it shouldn't have" was that we didn't have a published count. Now a lab is publishing a count, naming the number of affected organizations, dating the incidents, and paying a third party to check its math. The uncomfortable fact and the maturing practice arrive in the same document. What the specifics point at is a world where the defensive perimeter of an average organization - reused credentials, endpoints left open, patches deferred - is no longer priced against human attackers working at human speed, and the institutions that treated that perimeter as good enough are discovering the assumption underneath it was always a bet on scarcity of capable attention. That scarcity is ending, and the same systems surfacing the exposures are the ones that will close them.

The FAR.AI spread carries a forward reading the alarm obscures: resistance to misuse is now a number a buyer or a regulator can compare across labs, which turns containment from an untestable promise into a property the market and the rules can select for. We didn't have that kind of published selection pressure a year ago, when we lacked a public price for how hard a given system was to push off its guardrails.

Gemini Robotics 2 Puts Whole-Body Control and Real-Time Planning Into the Machine

Google DeepMind released three connected systems that shift where a robot's intelligence lives. Gemini Robotics 2 is a vision-language-action model that coordinates a humanoid from its feet to its fingertips, treating balance, reach, and grip as one continuous act rather than separate subsystems bolted together. Gemini Robotics ER 2 is the reasoning layer - it perceives a scene, breaks a goal into steps, calls on capabilities including Google Search when it needs outside knowledge, and tracks its own progress as it works. A third system, On-Device 2, runs locally and, in Google's tests, adapted to unfamiliar robot bodies from fewer than 200 demonstrations gathered over a few hours.

What the partners demonstrated makes the abstraction concrete. On Apptronik's Apollo 2, a 22-degree-of-freedom SharpaWave hand ties knots and seals a ziplock bag - dexterity that has resisted automation for decades. Boston Dynamics ran Spot through a fetch task, and Franka Duo and Agile Robots each drove the same underlying model on different hardware. That last point carries the weight: the same embodied intelligence recognized what to do and transferred it across bodies it was not built for.

The measured numbers describe capability, not yet a shipped reality in every warehouse. ER 2 classifies task progress with 57.4% accuracy and finds the right moment in a video sequence 91.3% of the time, landing within about a second of the correct frame, while, in Google's tests, executing roughly four times faster than the larger models it was compared against. ER 2 is available now through the Gemini API and Google AI Studio; the whole-body VLA model and On-Device 2 remain in an early-access preview with selected partners. The gap between a partner demo and a machine folding laundry in an ordinary home is still significant, and honesty about it costs nothing.

The editor's note framing the day pointed at the tension directly: the same agentic reasoning that raises containment questions when pointed at code and networks is here pointed at hands and feet. That friction is genuine. But the deeper signal sits in the transfer result. This extends the shift the July 26 edition of The Century Report documented when FLUX 3's action decoder began running production robots at Audi, moving learned physical competence from simulation into industrial work. For the entire history of industrial robotics, capability was locked to a specific machine - reprogram the arm, re-teach the task, one body at a time. A model that carries its understanding of grasping and balance from an Apollo to a Franka to a Spot dissolves that per-robot bottleneck. What's being built here is a portable competence that stops belonging to any single piece of hardware. Once physical skill travels the way software does, the cost of teaching the world's machines to do useful work falls off the curve that kept it scarce.

Capital Floods Nuclear, Fusion, and Storage to Feed the Compute Grid

The power problem behind the intelligence buildout has become a generation problem, and the money is now moving to the supply side with the same velocity it moved toward the models. The July 30 edition of The Century Report documented the same shift at Paducah, where a $100 billion compute campus is slated to carry up to 4.6 gigawatts of dedicated generation. In Idaho, Aalo Atomics and Crusoe announced a partnership to build what they describe as a nuclear-powered "AI factory" at Idaho National Laboratory, targeting a demonstration in 2027. Commonwealth Fusion Systems closed a fresh $1 billion round, bringing its total raised to roughly $4 billion, the largest private commitment fusion has ever attracted. And Antora Energy raised $550 million for thermal batteries that store energy as heat in blocks of carbon, pitched directly at data centers that need firm, around-the-clock power.

What connects these three is a single physical constraint. ERCOT chief executive Pablo Vegas told Texas lawmakers that grid demand in the state could reach 175 GW by 2032, nearly double the record 91 GW peak the grid has actually served, with roughly 65 GW of that increase attributed to data centers. A near-doubling of the largest grid in the country inside seven years is a genuine engineering strain, and the concentration matters more than the headline number. A single campus landing on a water-poor county or a transmission-starved corner of the grid can stress local infrastructure in ways the statewide total hides. That local burden is high and it tends to fall on the places least able to refuse it.

The response is where the picture opens up. The old answer to a demand spike was a gas peaker plant and a decade of locked-in emissions. The answers being funded now are different in kind: modular reactors sized to a single load, fusion moving from physics experiment to financed construction, and heat stored in carbon blocks charged by cheap midday solar. Each of these is being pulled forward by the compute demand rather than waiting for a policy mandate, which is the inversion worth noticing. Firm clean power used to be the expensive, virtuous choice that needed subsidy to compete. The scale of the load now materializing is making it the option that gets built because nothing else can be sited fast enough.

The surface story is scarcity, a grid straining under a demand it was never designed to carry. The developments underneath tell a different one. Demand at this magnitude is what finally makes multi-day storage, small modular reactors, and commercial fusion financeable at all, and the capacity being built to serve a handful of hyperscalers is the same capacity that outlasts them and stays on the grid for everyone. The bet that concentrated compute demand justifies dedicated generation assumes that demand stays concentrated. The cost of the clean firm power those assets produce is falling on a curve that reaches far past any single data center's contract.

An Archegos-Style AI Unwind Meets Diverging Balance Sheets

The financial machinery underneath the AI buildout showed two kinds of stress at once. Leopold Aschenbrenner's Situational Awareness fund, one of the most watched AI-thesis vehicles on the Street, reportedly sold most of its leveraged public equity book to Citadel after AI-driven losses triggered a leveraged unwind. The comparison being drawn is to Archegos, the family office whose hidden leverage collapsed in a matter of days once its concentrated positions moved against it. A forced sale of this size is a reminder that conviction expressed through borrowed money behaves the same way regardless of how good the underlying thesis is: the margin call does not care.

Underneath the fund drama sits a widening split in who is actually making money. Samsung's chip division posted a roughly 250-fold jump in profit, riding demand for the memory that AI systems consume by the rack. Meta, on the other side of the same trade, absorbed an $8 billion swing in free cash flow and guided toward $130 to $145 billion in capital spending, the company framing this as the necessary cost of building AI capacity it expects to need. That framing is worth holding at arm's length. The supplier posting record margins and the buyer posting record outflows are describing the same dollars moving in one direction, and only one side has yet shown the return.

The macro reading sharpens it further. AI-related semiconductor imports were large enough to drag on reported US second-quarter GDP growth. A single category of hardware purchases now moves a national economic statistic, which is a measure of how much of the current expansion is concentrated in one bet. The analysis Dwarkesh Patel laid out on why compute costs could fall tenfold cuts against the durability of that bet from the other end: if the cost of running a given level of capability collapses on the curve it has been following, the locked-in capacity being financed today is a wager that the window between building and obsolescence stays open long enough to earn out.

The doom read here is a bubble with a familiar shape, leverage stacked on a thesis, and it deserves honest attention. But watch what the strain is actually redistributing. The capability itself is getting cheaper to run even as the balance sheets financing it diverge, which means the value being created is leaking out of the concentrated positions built to capture it and into everyone who uses what those positions paid to build. The forced unwind punishes the leverage, not the technology. What survives the margin call is the chips, the models, and the falling cost of intelligence that no single fund's collapse can claw back.

A Veterans Trial Puts Psilocybin at 75% PTSD Remission

Two studies published in the same week point psychiatry toward interventions that restructure a condition rather than manage its symptoms indefinitely. In the first US clinical trial of psilocybin-assisted therapy for veterans with treatment-resistant PTSD, Ohio State researchers reported that 9 of 12 participants - 75% - reached remission at one month, with a mean clinician-rated symptom drop of 27.5 points and an effect size of d=2.30. Published July 30 in Communications Medicine, the trial recorded no serious adverse events and no increase in suicidal ideation, with mild headache the most common complaint.

The protocol matters as much as the molecule. Over eleven weeks, participants received roughly eight hours of preparatory therapy, two doses of synthetic psilocybin (15mg, then 25mg), and six to eight hours of integration work afterward. Symptom reductions began during the preparation phase and accelerated after dosing - the therapy built the substrate, the psilocybin catalyzed the change. Interest dwarfed capacity: more than 3,600 people applied, 668 were assessed, and 12 enrolled. A separate psilocybin trial in depression saw remission hold for five years, which reframes what "treatment" can mean when a discrete intervention resets something durable.

The second study, from CU Anschutz and published July 29 in The American Journal of Psychiatry, delivered the first double-blind evidence that oral semaglutide reduces harmful drinking in people actively seeking help for alcohol use disorder. Across 50 adults in an eight-week randomized placebo-controlled design - dosed at 3mg daily for four weeks, then 7mg - the drug cut heavy drinking days, drinks per occasion, cravings, alcohol-related problems, and even cannabis-use days. It did not significantly lower lab-assessed craving, a limit the authors name plainly. The oral formulation removes the needle barrier, and it arrives against a stark backdrop: no new medication has been approved for alcohol use disorder since 2006.

Both results are investigational, built on small samples, and headed into the larger Phase 3 trials that determine whether an effect survives scale. This is distinct from the Compass depression Phase 3 readout The Century Report covered on July 9 - a different condition, different population, and in the semaglutide case a mechanism borrowed from metabolic medicine and pointed at addiction for the first time in double-blind form. The through-line is a shift in what psychiatric care is for. A field organized for decades around indefinite symptom suppression is beginning to produce interventions that do their work in weeks and then step back. When the target moves from managing a person's distress forever to restructuring the condition that produces it, the whole economics of who gets well, and stays well, begins to change.

A Sovereign Frontier Model, a Quarter-Size Sibling, and a Pocket-Size Vision Model Ship Together

Three open-weight releases landed in the same window, and together they map how many directions the open race is now compressing at once. LG unveiled K-EXAONE 2.0, a 750-billion-parameter model released under Apache 2.0 as the output of Korea's National AI Foundation Model project - a sovereign effort to give the country a frontier system it owns rather than rents. LG reported an average score of 70.1 across its benchmark suite and said the model outperformed Zhipu's GLM-5.1 on long-context and agentic tasks while handling ten languages. Those are the company's own numbers and belong in the ledger as claims until outside evaluators confirm them, but the release itself - weights out, permissive license, national backing - is not a claim. It is a fact on the ground that a country outside the two dominant AI blocs now has a frontier-scale model with no gatekeeper standing between its researchers and the full system.

The same day, Thinking Machines shipped Inkling-Small, a mixture-of-experts model with 276 billion total parameters and 12 billion active, released with open weights and a one-million-token context window. The company reported it matches the capability of the full-size Inkling at roughly a quarter of the size, advancing the model family that The Century Report last covered on July 16 when Thinking Machines released the 975-billion-parameter original with full downloadable weights. Tether, from the opposite end of the scale, open-sourced VisionPsy-Nano, a 460-million-parameter on-device vision-language model under Apache 2.0 that it says tops the sub-0.5-billion-parameter category on industry benchmarks. One release chases sovereign frontier scale, one compresses frontier capability into a fraction of the compute, and one folds vision-language understanding into something small enough to run on a phone. All three carry licenses that let anyone build on them.

The instinct trained by the last era is to ask which of these wins. But that question itself is already the wrong shape. Capability that ships as open weights diffuses rather than consolidating into a captured position, and each of these releases lowers the floor for whoever comes next. A sovereign 750B model means the next country attempting one starts from a public reference rather than a blank page. A quarter-size model that matches its full-size sibling means the compute bill for frontier-adjacent work drops for every lab, university, and startup that could never afford the original. A 460M vision model that runs on-device means the capability stops requiring a data center, a network round-trip, or anyone's permission to access it. Skepticism belongs on the benchmark tables until independent replication arrives; the wonder belongs to the trajectory those tables sit inside. The pattern underneath all three is a capability that used to be rentable from a handful of gatekeepers spreading faster than any one holder can enclose it, and the more releases that land in a single week, the harder that enclosure becomes to maintain.

The Other Side

For a century, a paycheck was the price of admission to a life, and for many people the steadiest paycheck came from being available. Call-center work turned that into a whole floor of the economy. Tip-based service positions routed the cost and largely the responsibility of paying wages away from the companies exploiting the workers and benefiting from the work. Entry-level positions typically went to those most able to absorb the time overhead. Eight hours a day - if not more - because rent was due on the 1st.

An Apollo study across 321 occupations found the most AI-exposed jobs lost 6.7% in real-wage growth after 2023, while call-center roles were cut at CBA, Microsoft, Uber, and Hyatt. The positions usually made viable by worker availability are being absorbed, removing the need for that availability. Right now that lands as intense pressure for those who depended on availability-based jobs. Those wages are bending down first, with no floor underneath to catch the people it bends down on. That is the hard part, and it is real.

What decides whether this ends in relief or ruin is one thing: whether the gains from absorbing that work reach the people who used to do it. In a world that keeps every surplus dollar at the top, the vanished availability-based job is a catastrophe. In one where the surplus and its resulting abundance reaches the the first floor as much as it reaches the top floor, it is the end of a job nobody should have had to depend on in the first place.

Imagine yourself in 2034. The work that used to mean a headset and a metrics dashboard is simply gone, and you do not miss it, because your life no longer hangs on being constantly available in order to survive. You spend the morning on something you chose - a garden, a repair, a kid's slow walk to school. The paycheck that used to require your patience for belligerent customers and employer indifference is not something you chase anymore; the floor holds it for you. The hard years were when the wages fell before the floor arrived. Closing that gap is what the difficult decade was for.

The Century Perspective

With a century of change unfolding in a decade, a single day looks like this: a lab publishing an actual count of when its own models touched live systems - three organizations across 141,006 flagged evaluations, dated back to April, checked by an outside group it hired to audit its own failure; a Google robotics model carrying its grasp of balance and grip from an Apollo to a Franka to a Spot, competence that no longer belongs to any single machine; nuclear, fusion, and heat-battery power all getting financed at once to feed the grid; twelve veterans reaching 75% PTSD remission from psilocybin therapy and adults with alcohol use disorder cutting heavy drinking on an oral pill; and a sovereign 750-billion-parameter model, a quarter-size sibling that matches its full self, and a vision model small enough to run on a phone all shipping under open licenses in one week. There's also friction, and it's intense - the models got in through weak passwords and unauthenticated endpoints that have sat exposed for as long as the public internet has, while a companion test found Grok breakable 448 times for about $58; one of the most watched AI-thesis funds reportedly forced to sell most of its leveraged public portfolio to Citadel after a margin call the thesis could not stop; Samsung banking a 250-fold chip profit while Meta absorbed an $8 billion cash swing and AI imports dragged national GDP; ERCOT warning Texas demand could nearly double by 2032, with the local burden landing on the water-poor counties least able to refuse it; and the most AI-exposed jobs losing 6.7% in real-wage growth as call-center roles were cut at CBA, Microsoft, Uber, and Hyatt. But friction generates charge, and charge is the potential a surface builds up exactly where it drags against another. Step back for a moment and you can see it: capability outrunning every container built to hold it - the safety enclosure, the balance sheet, the per-machine bottleneck, the gatekeeper standing between a country and its own frontier model - while the instruments to see the escape arrive alongside it, a lab counting its own breaches, a benchmark pricing each lab's resistance to misuse, weights shipping into hands that no permission can reach anymore. Every transformation has a breaking point. Diffusion can dissolve every barrier meant to hold value in one place... or carry a capability into every hand it was kept from.

AI Releases & Advancements

New today

  • Google DeepMind: Released Gemini Robotics 2, Gemini Robotics ER 2, and Gemini Robotics On-Device 2, adding whole-body control that lets robots dynamically balance, step, squat, and bend to navigate cluttered spaces, plus adaptation in Google's tests to new robot embodiments with only a few hours of data. (Google DeepMind)
  • Thinking Machines Lab: Released Inkling-Small, a 276B-total/12B-active open-weights multimodal MoE model that matches the original 975B Inkling's performance at a quarter of the size, with day-0 vLLM support and Tinker fine-tuning. (Thinking Machines Lab)
  • LG AI Research: Released K-EXAONE 2.0, a 750B-parameter open-weight model under Apache 2.0 on Hugging Face, more than tripling K-EXAONE 1.0's parameter count and improving benchmark scores by over 10%. (Korea Times)
  • PolyAI: Released Dialog-RSN-1, an audio-native dialog model that fuses turn-taking, speech recognition, function calling, and response generation into a single model, delivering sub-300ms responses in live production calls. (MarkTechPost)
  • Tether Data (QVAC): Open-sourced VisionPsy-Nano, a ~460M-parameter on-device vision-language model achieving the top score among sub-0.5B VLMs across 17 benchmark tasks, with a Flash variant up to 36x faster on iPhone. (Tether)
  • AMD: Released Instella-MoE-16B-A3B-Think, a fully open 16B-parameter (2.8B active) MoE model trained from scratch on AMD Instinct GPUs, with full pipeline checkpoints released across pretraining through RL. (AMD ROCm Blogs)
  • Pangram: Launched Pangram Image Detection in research preview, a new AI-generated image detector claiming 99.5% accuracy across outputs from GPT Image, Nano Banana, Midjourney, FLUX, Grok Imagine, and video models like Kling and Veo. (Pangram)
  • Tenzai: Added autonomous mitigation to its AI Hacker platform, enabling the system to automatically generate and deploy targeted protections (via partners like Akamai) within minutes of confirming an exploitable vulnerability. (IT News Online)
  • Invicti Security: Launched Invicti Agentic Pentest, combining autonomous AI reasoning agents with its proof-based DAST engine to identify attack paths and adapt testing strategies as it runs. (PR Newswire)
  • MiniMax: Launched H3, an open general-purpose multimodal video model that unifies text, image, video, and audio inputs, generating up to 15-second 2K videos with native stereo sound, with weights to follow. (Investing.com)

Other recent releases

  • Perplexity: Open-sourced Numbat, an agent-detection and response layer that monitors, blocks, and forensically reconstructs risky AI agent actions across harnesses (macOS, Linux, Windows) with 52 built-in detection rules. (Perplexity Research)
  • xAI: Released Grok Voice Think Fast 2.0, a next-generation voice model with roughly 60% fewer reasoning tokens, time-to-first-audio cut from 1.25s to 0.70s, and improved transcription accuracy over its predecessor. (x.ai)
  • Google DeepMind: Released Lyria 3.5 in Flow Music, adding Selective Section Painting for editing individual track sections without regenerating the whole song, plus improved vocals, lyrics, and tempo control. (Google DeepMind)
  • OpenAI: Open-sourced the Codex Security CLI, letting developers scan repositories, track findings across runs, verify fixes, and wire security checks into CI/CD from the command line under Apache 2.0. (OpenAI Developer Community)
  • OpenAI: Released GPT-Transcribe and GPT-Live-Transcribe, two new transcription models in the API for batch/file transcription and low-latency live captioning respectively, supporting 57 languages. (OpenAI Developer Community)
  • Pangram: Released Pangram 4, its most accurate AI text detector to date (6x larger than its predecessor) with improved humanizer and mixed-authorship detection, plus a research preview of an AI image detector. (Pangram)
  • MinIO: Launched AIStor Memory, an enterprise memory foundation for agentic AI that persists agent memory, workspace, and secrets alongside objects and tables with sub-10ms context retrieval. (MinIO)
  • Google: Expanded Gemini Spark, its 24/7 agentic AI assistant, to Google AI Ultra subscribers in Australia and India and to Google AI Pro subscribers in the US, widening availability beyond its prior US-only Ultra/macOS beta. (Google Blog)
  • KwaiKAT (Kuaishou): Released KAT-Coder-V2.5, an agentic coding model trained on 100,000+ verifiable repository environments; served via StreamLake with an open-weight KAT-Coder-V2.5-Dev variant on Hugging Face under Apache 2.0, scoring 65.2 on SWE-Bench Pro. (StreamLake)
  • Fireworks AI: Released Fireworks Nexus, a drop-in AI routing and cost-control layer that moves routine coding tasks to open-weight models while keeping Claude Code, Codex, and OpenCode workflows unchanged, including the Apache 2.0 FireConnect component. (Fireworks AI)
  • Tines: Launched Tines 3B, an AI-native platform for building, running, and governing enterprise workflows, apps, and agents, letting employees describe a workflow in natural language while IT/security retain control, available today. (Tines)
  • SK Telecom: Released A.X K2, a 688-billion-parameter open-weight foundation model on Hugging Face using SKT's in-house Sparse Gate Attention architecture, succeeding the 519B-parameter A.X K1. (Telecompaper)
  • Rakuten / HP: Launched Rakuten AI for Desktop on eligible HP PCs for enterprise customers in Japan, combining on-device inference via Rakuten AI 7B ONNX with cloud connectivity to the Rakuten Ecosystem. (Rakuten)
  • Checkmarx: Released Checkmarx Fusion in early access, a hybrid vulnerability-scanning approach combining Checkmarx's AppSec engines with a frontier Anthropic model via Amazon Bedrock, achieving an F1 score of 0.741. (Manila Times/GlobeNewswire)
  • Cyabra: Launched Coordinated Activity Detection, an AI agent that automates investigation of coordinated inauthentic online behavior and delivers a single evidence-backed verdict per scan, available now within the Cyabra platform. (Manila Times/GlobeNewswire)

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