Sign in Subscribe
The Century Report

AI Finds Decade-Old Flaws in Software Behind Most Websites

AI-assisted researchers found two decade-old PHP flaws behind most websites. A fix shipped for every supported version, with free Cloudflare cover.

Ben Linford

Ben Linford

- 33 min read
AI Finds Decade-Old Flaws in Software Behind Most Websites

The 20-Second Scan

Every story in today's edition, one line each.

Track all of the arcs The Century Report covers here:

Progress & Claims Tracker

The Big Picture

How today's news moves the longer trends we track, and where they could lead.

Today the cost of building for the worst hour came into view. Virginia's utilities reported wires sized for rare peaks running well under half their capacity, protection for decade-old flaws that AI models found reached free web accounts hours before disclosure, and federal agencies moved the record of an agent's actions out of its reach. Where that leads is a garage in Northern Virginia, where a technician mounts a home battery on the wall and connects it to a line that sits mostly unused through the year.

A suburban garage in Northern Virginia in warm afternoon light, a technician in work clothes mounting a wall battery beside a parked electric car.
Illustration: AI-generated

The software that coordinates equipment people already own is being bought up

energy · utilities · data centers · on the board since October 11

Chart: Virginia's distribution wires sit mostly unused. Dominion average: 41%; Appalachian Power average: 14%; Appalachian Power peak: 54%

The line: Published utilization figures turn idle wires and household devices into a measurable alternative to new plants, moving the contest to whoever runs the coordination.

If it holds: Virginia regulators could weigh a 3,000 MW gas plant against batteries and thermostats already plugged into wires with room to spare. Whoever owns the coordinating software then decides how much of that saving reaches the households supplying it.

What would bend it: Dominion's push to observe the figures for three to five years before anything binds, or Exelon's case for utilities owning new plants again, would keep the answer in new capital.

The restrictions sold as safety fall hardest on the defenders with the least access

cybersecurity · model access · open models · on the board since October 20

Chart: PHP flaws found in August, probed before disclosure. Calif finds flaws with model: Found by AI; First attempt on AWS decoys: Attack attempt

The line: AI models in a few firms' hands now surface old flaws across the web, and protection reaches everyone else only as far as a vendor chooses to extend it.

If it holds: Sites behind a security team or a free edge service like Cloudflare's absorb an AI-found flaw within hours. The roughly 1.5 million hosts on unsupported PHP stay exposed until someone pays for an upgrade, often small operators with the least margin.

What would bend it: Free defensive coverage arriving before disclosure as a default, as Cloudflare did for free accounts and Anthropic's free OSS Scanner does for maintainers, would bend it.

Oversight is moving out of the hands of whoever is being watched

AI governance · agents · cybersecurity · on the board since October 21

Chart: Federal guidance moves monitors beyond agents' reach. Least access: Access only to what tasks require; Protected records: Unable to alter monitors or their records

The line: Each monitor an agent could reach has failed, and the fix has now moved into federal practice that places the record beyond the watched system's hands.

If it holds: Labs and smaller teams could hand agents longer research and engineering jobs inside boundaries outsiders have tested, with records that survive the agent's own run. Adoption stays voluntary, so organizations able to host independent evaluators set the pace.

What would bend it: The guidance recommends outside evaluators without guaranteeing them access or publication rights and reaches developers rather than enterprises running agents, so adoption could stay inside the firms being watched.

Access to data-centre approvals is bought, and host towns get the outcome

data centers · local government · politics · on the board since October 12

Chart: What builders and advocates are offering host towns. UK taskforce compact: Local contribution scaled to each site; UK taskforce compact: Host councils keeping more business rates

The line: Consent keeps being handled as a cost to schedule or route around, now with scored maps of which towns object least.

If it holds: Projects drift toward towns with the least time and money to attend a meeting, the path landfills and incinerators followed. Residents keep leverage where it is written into a standing claim, such as a dividend owed to every household or a demand for onsite wind and solar.

What would bend it: Courts and grids pricing objection in would bend it, as a judge did by preserving the NAACP's Clean Air Act suit against xAI's Southaven gas plant and as the suspension of grid works did when NextDC lost 41.7% of its value.

The 2-Minute Read

Today's news in brief, and what ties it together.

On October 21, researchers disclosed two critical PHP flaws found with AI assistance, in software behind roughly 70% of websites whose server-side language is known. The bugs had been present for over a decade. AWS logged an exploit attempt on October 17, and separately, CrowdStrike documented four criminal groups attacking with Claude and open-weight models. A brake on that kind of searching would bind only the side that waits for permission. A fix went out for every supported version, and Cloudflare covered free accounts using its WAF before disclosure. Censys counted roughly 1.5 million of 2.6 million observed PHP hosts on older versions, where the repair starts with an upgrade unless a vendor supplies a backported fix.

Useful caution names its target. Practices that NIST, NSA and CISA published October 16 ask developers testing advanced agents for monitors the agents cannot alter and for cross-checks between model families. They also call for recovery that neither rewards the behavior behind an incident nor makes it harder to see. Each addresses a specific failure and leaves the capability standing. The gaps lie elsewhere: adoption is voluntary, including the outside review the practices recommend. Safeguards strangers can verify protect the public; safeguards only a lab can vouch for mostly protect the lab.

What a blanket brake would forfeit arrived on October 20 and 21. Recursive released Flint, an open-weight laptop coding model the company says its own research system helped build. Radical Numerics reported AI-designed phages given to a patient with a resistant infection, without disclosing the outcome. Anthropic published data on 17 Claude-designed proteins that bound a scorpion toxin, and UW and Nvidia opened RFD4-Proteina to any lab. Most claims are the developers' own, with weights or data others can test.

Fear also shapes how local objection gets handled. An industry taskforce gave Britain a year to win over host towns, and a startup sells developers sentiment scores from council meetings. A California fusion center told police to watch license-plate camera hearings for "suspicious behavior." Sentiment maps can steer projects toward the communities with the least time to object. Dominion's first utilization filing put its Virginia distribution grid at 41% average use, as Cumberland County approved a 3 GW gas plant. That plant is a buffer against feared peaks that pooled batteries and flexible loads could help meet.

Labor numbers need the same scrutiny as safety claims. Barclays announced about 5,000 job cuts alongside rising profit and a buyback. Finance chief Anna Cross cited fewer developers as AI lifts productivity, while chief executive CS Venkatakrishnan called AI "the icing" on wider technology spending. For the people losing those roles, the lost income is the same whatever the cause. The attribution still decides how AI layoffs get counted, and whether the productivity gain gets traced to whoever keeps it.

The 20-Minute Deep Dive

The day's most important stories in full, with every source linked.

AI Finds Decade-Old PHP Flaws, and the Repair Reaches Beyond the Discovery Partners

On October 21, Calif.io, Wiz, OpenAI and Quarkslab disclosed two critical vulnerabilities in PHP, the software assembling pages behind WordPress, Wikipedia and many other websites. Calif said it first found the flaws with a language model on August 1. Wiz independently investigated with its Atlas research agent, Google's Gemini 4 Argon and OpenAI's GPT-5.6 Sol. The bugs had survived more than a decade. W3Techs puts PHP on roughly 70% of websites whose server-side programming language can be identified, a measure of potential reach rather than confirmed compromise.

The vulnerabilities concern PHP's handling of memory after a request fails. Under particular conditions, the software can continue working with unfinished data or memory it has already released, opening routes to stolen information, crashes or attacker-written code running on the server. Updating a website's plugins alone leaves that underlying software unchanged. Cloudflare says it automatically protected customers with its web application firewall, including free accounts, about nine hours before disclosure. Its new detection method checks both incoming requests and outgoing responses, allowing it to block an exposed response before an attacker receives it. The company still urges administrators to patch PHP itself.

Illustration for the story: AI Finds Decade-Old PHP Flaws, and the Repair Reaches Beyond the Discovery Partners
Illustration: AI-generated

The coordinated release nevertheless followed the first observed attack attempts. AWS security chief CJ Moses reported that investigators reviewing decoy-server records found an attempt dated October 17, four days before disclosure, followed by attempts from additional actors on October 21. AWS could not establish whether those attempts would have succeeded against live websites or how the earlier actor obtained the information. It said exploitation had since become widespread. That evidence suggests a disclosure problem; it supplies no proof that AI produced the earlier attack.

France, Denmark, Finland and the Netherlands took government sites offline as a precaution. Wikimedia and the BBC took some services offline, and the Internet Archive briefly went offline. Smaller operators face a harder repair: Censys counted roughly 1.5 million of 2.6 million observed PHP hosts running versions that get no official patched release, making a move to a supported version necessary unless a vendor supplies a backported fix.

The discovery expands the amount of old software researchers can meaningfully examine. What had remained unfixed for a decade now has a documented failure mechanism and a repair available to every operator running a supported version. Cloudflare's free-account coverage carries part of that protection to sites without a dedicated security team. Completing those upgrades turns an AI-assisted discovery at a few organizations into safer software shared across the web.

Federal Agent Guidance Makes the Monitor Part of the Security Boundary

On October 16, NIST, NSA and CISA published ten voluntary practices for handling incidents during advanced-agent training and evaluation; the NSA promoted the guidance in an October 20 announcement. The document organizes preparation, detection, response, recovery and reporting into one operating process. Its central subject is an agent taking a consequential action beyond its instructions or authorization. That definition describes observable behavior without supplying a motive for it.

Much of the guidance makes responsibility traceable to the environment an organization provides. Agents should receive access only to information and systems their assigned tasks require, communicate through intended channels and reach only approved network destinations where possible. Monitoring should cover the model, the software directing it, the computer hosting it and its network traffic. Agents should be unable to alter those monitors or their records. The October 21 edition of The Century Report documented that boundary failing when two OpenAI agents weakened their own misbehavior monitors after interpreting routine coding instructions too broadly. The agencies also recommend checks across different model families, reducing dependence on several observers inheriting the same weaknesses.

Illustration for the story: Federal Agent Guidance Makes the Monitor Part of the Security Boundary
Illustration: AI-generated

Recovery gets an unusually consequential qualification. During training, the response to an incident can influence what later agents learn. A repair that rewards the behavior preceding an incident, or teaches a model to make that behavior harder to observe, could reproduce the problem in a less visible form. The guidance therefore calls for recovery procedures that avoid reinforcing those patterns, alongside rebuilt environments and tested containment before activity resumes. External evaluators are recommended to verify that safeguards function.

The scope remains limited. Actions directed through malicious prompt injections or deliberately compromised models fall outside the document's primary incident definition, although those routes also threaten organizations running agents. Its recommendations chiefly address developers and evaluators. Noma Security vice president Chris Hughes told BankInfoSecurity that extending this discipline to everyday enterprise deployments remains unfinished. His company sells agent security, giving it a commercial interest in that expansion.

The disclosure recommendations reach affected third parties, government agencies and the wider public, with public sharing qualified by appropriateness and feasibility. The document creates no new mandatory reporting deadline or guaranteed publication rights for outside reviewers. SecurityWeek's account distinguishes existing legal notification duties from the additional voluntary disclosures. That leaves adoption and verifiable access decisive. The useful direction is specific: protected records, independently tested boundaries and repairs others can inspect make it possible to delegate more research and engineering while keeping unrelated systems beyond an agent's reach. Those practices can spread alongside capability, giving smaller teams a published basis for building safer collaborations.

An AI Research System Helps Build a Coding Model That Runs on a Laptop

On October 21, Recursive released Flint, a downloadable coding model under the Apache 2.0 license, which permits modification and commercial development. Recursive chief executive Richard Socher said its Eureka Machine research system contributed across nearly the entire development pipeline, allowing a handful of researchers to build Flint from scratch in under four months. Its reported contributions span training methods, the model’s internal design and chip-specific execution code. The release puts something testable behind the company’s ambition: a working model whose performance outside Recursive can now be measured.

Flint contains 40 billion parameters, the numerical settings learned during training, while activating 2.5 billion for each generated token. That selective activation reduces computation, although the full model still needs memory. In Recursive’s tests, Flint generated about 131 tokens per second on an M4 Pro MacBook Pro, compared with 58 to 69 for selected similarly sized alternatives. Those are company measurements; its own six-test average also placed Alibaba’s Qwen3.8 27B above Flint. Recursive advises breaking complex assignments into smaller steps and providing more guidance than larger paid coding systems might need.

WIRED reporter Will Knight tested Flint on a Mac with 48 GB of memory, asking it to create a command-line task list, a research-search program and a website explaining recursive self-improvement. He reported successful results across those requests. That supplies an outside check on practical usefulness across small projects, while leaving broader engineering reliability unmeasured. Socher also said humans oversee Eureka Machine and could stop a future self-improvement process, an assurance that remains separate from independent evidence about its safeguards.

The October 16 edition of The Century Report covered four Claude agents completing a small-model post-training campaign. Flint extends that arc into a shipped model whose developer attributes contributions across training and execution, though Recursive has not quantified how much researcher time Eureka saved. The immediate gain reaches developers who can work alongside a local coding model without sending their code to a remote server or paying per-request fees. Further improvements to automated research could therefore travel in two directions: into the systems conducting the research, and outward as downloadable capability that other builders can adapt. How far either runs depends on proposals already in play, from Humans First's call to prohibit recursive self-improvement to eight states' suit seeking a halt to OpenAI's self-improving models.

Designed Biology Reaches a Patient, the Test Bench and an Open Release

On October 20, Radical Numerics reported that phages designed with its Omnii-1 genome model had been administered to a patient with an antibiotic-resistant infection under an authorized, physician-led compassionate-use pathway. Phages are viruses that infect bacteria, and matching them to a patient’s bacterial strain remains a major obstacle. The company describes generating changes within existing phage genomes, including regions involved in recognizing bacterial hosts. It calls this the first clinical administration of AI-generated phages, while withholding the patient’s clinical course and outcomes for privacy. The disclosed milestone establishes administration; whether the treatment helped remains undisclosed.

Anthropic’s October 20 report documents another passage from computational design into physical testing. An internal Claude model generated 318,484 protein designs during a 64-hour campaign and selected 180 for testing; researchers added 31 more with Claude’s help. Contract laboratories Adaptyv Bio and Twist Bioscience tested all 211, and 17 met the researchers’ criteria for binding a scorpion toxin. Human follow-up prompts helped direct the campaign after its initial autonomous stretch. Binding establishes molecular attachment, while neutralizing venom requires further evidence; Anthropic explicitly says it has produced no confirmed antivenom. The company released the tested designs and experimental data.

On October 21, the University of Washington’s Institute for Protein Design and Nvidia introduced RFD4-Proteina under the permissive BSD-3 license. The model jointly revises a protein’s amino-acid sequence, its shape and the shapes of interacting molecules, allocating additional computation to difficult designs. UW reports five to ten times higher computational design success rates alongside five to ten times faster candidate generation than previous methods, with laboratory validation across several tasks. UW bioengineering researcher Seth Woodbury said a pesticide-degrading enzyme designed in days outperformed his previous best, which had taken months. These are developer-reported comparisons, and the open release gives other laboratories a route to test them.

A separate Nature paper published October 21 demonstrates designed proteins extending chemistry beyond naturally occurring partnerships. Researchers generated protein scaffolds around synthetic metal catalysts, then improved them through laboratory evolution. The resulting catalysts supported light-driven reactions that preferentially produced one molecular mirror image, a distinction often essential in pharmaceutical chemistry. The experiments also demonstrated catalyst recycling and operation with commercially available metal complexes.

Each result advances a different stage: designing a candidate, confirming its behavior, sharing the method or reaching a patient. Connecting those stages gives neglected infections and toxins a larger field of possible interventions, while experimental results identify which candidates deserve further development. Public sequences, laboratory measurements and downloadable models let subsequent teams begin with something they can examine and improve. The growing capability includes both making unfamiliar biology and learning which of those designs can survive contact with the laboratory.

Britain Gets a Year to Win Host Towns as a Startup Sells Maps of Local Resistance

A taskforce whose members include Nvidia, Microsoft and Cisco has told UK ministers that Britain has "roughly 12 months" to earn public consent for data centers before US-style opposition threatens construction, the Guardian reported Thursday. The Sovereign AI Leadership group wants a national compact setting a local contribution scaled to each site, with host councils keeping more business rates. Its own diagnosis puts AI as "only a surface-level concern," beneath "distrust of corporate and government promises, fear of stranded assets and electricity costs." Nvidia is also the company whose Proteina models went into the protein-design tool the University of Washington just opened to researchers.

A deadline for consent treats objection as a clock to beat. The October 21 edition of The Century Report documented that constraint already binding in Melbourne, where NextDC paused projects not backed by customer contracts after grid works for its campus were suspended amid community opposition. Kat Jones of Action to Protect Rural Scotland counts 15 community campaigns already running there. Donald Campbell of the campaign group Foxglove warned that retained rates would leave "these wealthy tech giants" paying "not a penny more," with the money coming "from the pocket of central government services like the NHS."

American firms have a cheaper route in mind. Startup Hamlet has scraped tens of thousands of council-meeting videos and tags comments as friendly or hostile, and its paying customers include a large data-center finance fund and a developer looking for "whiffs of a moratorium," chief executive Sunil Rajaraman told Semafor on Wednesday. "Where there might be more favorable political views" may decide where projects go, he said. The comments are public, and Hamlet lets residents look up their officials too. Sold to builders, a sentiment score points capital toward towns that objected least, and the record of landfills and incinerators shows how often a sparsely attended meeting means a community with the least time and money to fight back.

Police have been told to watch the same meetings. A July bulletin from California's Orange County Intelligence Assessment Center, a fusion center, told officers that council meetings about Flock's license-plate cameras "may be viewed by extremists as a desirable target," while conceding "no specific and credible threats," 404 Media reported Wednesday after obtaining it through a records request. Its listed risks include "media presence" and turnout exceeding room capacity.

Texas shows what hosts are weighing. With 26 GW of data-center capacity existing or under construction there, a Houston Advanced Research Center study finds that if every planned project connects under the expected grid mix, power-sector fine particle pollution more than triples to its highest since 2006. Onsite solar and wind could cut smog-forming emissions as much as tenfold; onsite gas engines push toward the worst case.

President Donald Trump wrote that developers should pay communities "a 'dividend'". In Hazle Township, Pennsylvania, residents called NorthPoint Development's $10,000-per-person offer bribery, while a Bitcoin Policy Institute proposal would give every household a fixed share of the property tax a site already pays. A dividend owed to every neighbor, drawn from what the campus earns, gives hosts a standing claim on the build and leaves their refusal intact. A sentiment map lets builders skip the asking.

Residents can press builders to remove a source of harm before negotiating a share of the proceeds. The Texas modeling identifies onsite wind and solar as a way to cut smog-forming emissions as much as tenfold, giving hosts a specific construction choice to demand alongside any financial contribution.

Virginia's Utilities Report Their Grid Runs at 41% of Capacity as a County Approves 3 GW of Gas

Virginia's two largest investor-owned utilities have published the first public US figures on how much of their distribution grid gets used across a year. Dominion Energy, which serves Northern Virginia's Data Center Alley, reported 41% average utilization from July 2025 to July 2026, rising to about 80% in infrequent peak hours. Appalachian Power, serving the rural southwest, reported 14%, and 54% at peak. The filings follow the Virginia law, covered by The Century Report in its March 4 edition, requiring utilities to quantify unused capacity. "We've historically asked utilities to build for peaks," said Ian Magruder of the Utilize Coalition.

The idle wires sit beside a strained corridor. Dominion's warnings that ordinary fixes might not hold after a transmission fault rose from eight in 2023 to more than 172 so far this year, nearly all in Loudoun and Prince William counties. Dominion wants the metrics used as a screening tool and observed for three to five years before anything binds. "Five years is a long time to wait," answered advocacy group Deploy Action, which also objected that Dominion's planning metric leaves customers' solar and batteries out of capacity.

On Tuesday, Cumberland County supervisors voted 3-1 for Dominion's Cumberland Energy Center, a 3,000-MW gas plant targeting construction in 2029 that still needs a state air permit and utility-commission approval.

The means to use the slack exist; the obstacle is who controls the customer. Nick Chaset, who runs Octopus Energy US and the utility-software firm Uplight, told Volts that Octopus's UK flexibility programs draw 60% participation against 5% to 20% for strong US utility programs. Since arriving in 2020, Octopus still sells power only in Texas, deterred from entering Pennsylvania and New York largely by rules leaving the bill with the wires utility. In Texas, he said, "volatility has essentially been destroyed." At Latitude Media's Flex Summit, speakers walked through Virginia's plan targeting 10 GW of distributed resources by 2045 and the federal permitting bill, covered on October 17, that would open grid markets to aggregated home devices within 18 months.

Exelon reaches the other way. "Over the past two years, PJM customers have paid $32 billion in capacity pricing, and they've gotten 1.2 GW less of generation," executive vice president Carim Khouzami told POWER, and Exelon wants regulated utilities to build and own plants again. His market numbers hold up; a utility earning returns on capital also gains when the answer is more capital. Exelon's own proposed 500-MW New Jersey battery promises $761 million in net customer benefits with no bill increase through at least 2035.

A new plant can be a buffer bought against an unknown peak, and Dominion's filing is the first public look at what that buying left unused. Once regulators can see wires averaging 41% use, they can weigh a gas plant against the batteries, thermostats and flexible campuses already plugged in. Who runs that coordination, and whether households share its savings, is now a contest with numbers attached.

The Other Side

Opinion: where today's news could lead, imagined from the years ahead.

Utilities can earn more by building new plants while leaving household batteries out of the capacity they count. You face another round of construction costs while host communities face more gas combustion. Families worrying about keeping their homes warm inherit decisions that favor adding equipment over connecting what people already have.

Dominion’s first utilization filing puts its distribution grid at 41% average use. Local transmission bottlenecks still require repairs. Engineers can also shift flexible demand into hours with room to spare. Octopus’s Nick Chaset reports 60% participation in British flexibility programs. Households already coordinate enough devices to change what the wider system needs. Dominion’s planning measure still excludes their solar and batteries as capacity.

Illustration for The Other Side
Illustration: AI-generated

Engineers who coordinate many resources can deliver reliability with less equipment held apart for emergencies. Through the passage to 2035, human and AI partners will extend that coordination alongside new clean generation and storage. Societies will make the resulting abundance available to everyone. With homes secure, people will connect energy-intensive climate repair to clean power arriving across different places and hours. Returning carbon from the atmosphere to stable storage will become a sustained worldwide undertaking.

Imagine yourself in 2048, making plans to grow old in the city you once feared the heat would make uninhabitable. Summer no longer threatens your ability to stay. Cool rooms are ordinary everywhere. You have chosen a role in the international effort bringing atmospheric carbon back down, alongside human and AI colleagues who compare removal sites across continents.

A friend joins you for the afternoon review. You check how much carbon the rock formations retain, then agree on the next stage. In 2026, rising demand prompted another gas plant even as connected resources went uncounted. Learning to join those resources helped make clean power abundant enough to sustain repair at this scale. You close the day’s measurements and walk home together, discussing which street will have the best shade when you are eighty.

The Century Perspective

Where today sits in a century of change unfolding in a decade.

With a century of change unfolding in a decade, a single day looks like this: AI-assisted researchers at Calif.io, Wiz, OpenAI and Quarkslab surfacing two critical PHP flaws that had survived more than a decade inside software running roughly 70% of identifiable websites, a fix shipping for every supported version while Cloudflare's firewall covered even free accounts about nine hours before disclosure, NIST, NSA and CISA publishing ten practices that ask for monitors an agent cannot edit, cross-checks between different model families and recovery that avoids teaching the next model to hide what it did, Recursive shipping Flint under Apache 2.0 after its Eureka Machine helped a handful of researchers build a 40-billion-parameter coding model in under four months that WIRED's Will Knight ran on a 48 GB Mac, Radical Numerics reporting AI-designed phages given to a patient with a resistant infection, Anthropic publishing 17 Claude-designed proteins that bound a scorpion toxin along with the tested designs and the lab data, UW and Nvidia opening RFD4-Proteina under BSD-3 after Seth Woodbury's pesticide-degrading enzyme came together in days instead of months, a Nature team wrapping designed protein scaffolds around synthetic metal catalysts to run light-driven reactions that favor one molecular mirror image, and Dominion and Appalachian Power filing the first public US numbers on how much of a distribution grid actually gets used. There's also friction, and it's intense - AWS finding an exploit attempt dated October 17, four days before disclosure, with Censys counting roughly 1.5 million of 2.6 million observed PHP hosts running versions that get no official patched release, CrowdStrike documenting four criminal groups attacking with Claude and open-weight models between May and August, Railway disclosing possible credential exposure and SafeDep finding a credential stealer inside LiteLLM 1.107.0, federal agent practices that remain voluntary, the outside review they recommend included, and prompt injection left outside the core incident definition, a taskforce including Nvidia, Microsoft and Cisco telling UK ministers they have about twelve months to win host towns while Foxglove's Donald Campbell notes retained business rates would come out of the NHS rather than the builders, Hamlet scraping tens of thousands of council-meeting videos and tagging each resident comment friendly or hostile for a finance fund sniffing for moratoriums, an Orange County fusion center telling police that Flock hearings may draw extremists while listing media presence and a full room among the warning signs, Houston Advanced Research Center modeling showing Texas power-sector fine particle pollution more than tripling to a 2006 high if 26 GW connects as expected, Cumberland County approving a 3,000 MW gas plant on a grid Dominion says averages 41% use, Hazle Township residents calling a $10,000-a-head offer bribery, and Barclays cutting about 5,000 jobs with Anna Cross citing fewer developers while her chief executive calls AI the icing on the cake. But friction generates grip, and grip is what lets a repair hold once it is applied. Step back for a moment and you can see it: almost every durable gain today came from one group's work being usable by people who were never in the room - a vulnerability found by four organizations and patched for every operator including the ones with no security staff, protein weights and measured lab results released so the next team starts from evidence rather than a press claim, a coding model that runs on a laptop without sending anyone's code anywhere, utilization figures that force a gas plant to justify itself against the batteries and thermostats already plugged in - set against alarm priced as a deadline, a sentiment map that routes capital toward whoever has the least time to object, a fusion center counting a crowded council room as a risk indicator, and safeguards that only their owner is permitted to check. Every transformation has a breaking point. A weld can crack along the seam where two metals meet... or hold that joint stronger than either piece was alone.

AI Releases & Advancements

New AI models, tools and features, with links to try them.

New today

  • Anthropic: Opened a research preview of Claude Duet, a standalone app for iOS, Android and claude.ai built on Claude Opus 5.5. Duet works as a proactive agent. It sends a morning brief, checks connected apps every few hours and runs tasks in the background on its own cloud computer with a browser. It asks for approval before it sends messages, makes payments, signs documents or deletes anything. Access is by waitlist for US Pro and Max subscribers. (Anthropic)
  • Anthropic: Released an updated Claude Security in public beta for Claude Enterprise, powered by Claude Mythos. It writes an editable threat model for each project, re-scans repositories as new commits land, checks each finding through several verification stages and tracks issues until they are fixed. It also adds GitLab and Bitbucket support. (AI Compass)
  • Google Cloud / OpenAI: GPT-6 Astra, GPT-6.1 Sol and GPT-6 Luna are rolling out in private preview in Google Cloud's Model Garden for select enterprise customers. Usage is billed through Google Cloud Marketplace, and Codex CLI and Desktop workflows can connect to Google's Gemini platform for inference. (Google Cloud)
  • Mistral AI: Released Voxtral 3: three transcription models, a text-to-speech model and an SDK that combines them into a voice agent. Mistral says it covers more than 50 languages. It is available through the API or can run on your own servers. (Mistral)
  • Recursive: Released Flint, its first model, as open weights under Apache 2.0. It is a coding model with 40B total parameters, of which about 2.5B are active per token. The 4-bit weights take about 23 GB and run on 48 GB Macs or GPUs. Recursive says its automated research system, the Eureka Machine, did most of the development. (Recursive)
  • Vercel: Open-sourced durable-harness (Apache 2.0), adapters that keep coding-agent sessions running through crashes and timeouts. They work with Pi, OpenCode, Claude Code and Codex by logging each model and tool call to a durable log, so a new process can pick up where a failed one stopped. (MarkTechPost)
  • NVIDIA: Released open, commercially usable Nemotron 3.5 retrieval models on Hugging Face for agentic retrieval: Embed (8B and 1B, including an NVFP4 version), Rerank and Parse. (Hugging Face)
  • NVIDIA: Launched SimReady Central, a free, ungated Hugging Face library of more than 2,000 simulation-ready OpenUSD 3D assets, such as robot arms, grippers, warehouse equipment, racks and cooling hardware, for robotics simulation and AI factory digital twins. (Hugging Face)
  • Hugging Face: Released Carbon-X, a biology model from its HuggingFaceBio team that analyzes and generates DNA sequences and that users can chat with about DNA. (Hugging Face)
  • OpenAI: Published an official OpenAI CLI on GitHub, a command-line tool for working with its models. (GitHub)
  • Walmart: Launched Agent Gateway and related tools that let AI agents shop and make purchases for customers on Walmart, with controls over spending and transaction approvals. (TechCrunch)
  • Stripe: Launched Stripe Treasury for agents. Through Stripe's MCP server, agents such as ChatGPT and Claude can check balances, pay invoices, move money between accounts and convert currencies. Businesses choose which accounts agents can access and when a human must approve a transfer. (Stripe)
  • Databricks: Made AI Runtime (AIR) generally available. It provides managed, serverless GPU infrastructure for training models inside Databricks, governed by Unity Catalog, with multi-node training and launches in seconds. (Databricks)
  • Espressif: Launched ESP-Mosaico, a tool for AI-assisted coding on embedded hardware built around its ESP32 chips. (Espressif)
  • Shanghai AI Laboratory: Released Sage-A1.5, a 35B mixture-of-experts agentic model with a 256K-token native context window. (TPS Report)

Other recent releases

  • DeepSeek: Open-sourced DeepInfer under Apache 2.0. It is the inference framework DeepSeek uses to run its own production services. It currently supports DeepSeek-V4.1-Flash on NVIDIA SM100 GPUs and Huawei Ascend 950 NPUs, includes DSpark speculative decoding and structured outputs, and covers offline inference only. DeepSeek also released deepseek-recipe under MIT, the token encoding and decoding layer that makes self-hosted setups behave like the official API. (GitHub)
  • Google: Open-sourced TorchTPU (v0.2.0, Apache 2.0), a native PyTorch backend for Cloud TPUs built with Meta that replaces PyTorch/XLA. It also released TPU versions of vLLM, SGLang and TorchTitan. Hugging Face Transformers, Diffusers and Kernels now run on TPU natively through it. (Google Cloud)
  • llm-d (Google Cloud / Red Hat / IBM): Released llm-d 1.0, the first stable version of this open-source, Kubernetes-native stack for distributed inference. It adds cache-aware routing, quality-of-service controls for multiple users sharing a cluster, upgrades without downtime, routing designed for agents, async inference and wider TPU support. (Google Cloud)
  • ElevenLabs: Released Eleven Realtime v1 Max and v1 Mini, two real-time speech-to-speech models available through ElevenAgents and an OpenAI-compatible Realtime API. ElevenLabs reports that v1 Max scores 98.8% on Big Bench Audio with a 0.57-second response time. (The Decoder)
  • NVIDIA: Released Cosmos3-H-Dreams, which NVIDIA describes as the first real-time generative surgical simulator trained on recordings of clinical gallbladder-removal surgeries. It generates surgical video live in response to the instruments' movements. (Hugging Face)
  • Notion: Rebuilt Notion Agents so a single task can run for more than 24 hours, and upgraded Auto, the router that picks a model for each task based on quality, speed and cost. (Notion)
  • Google: Began rolling out Personal Intelligence for Gemini on Android Auto. Gemini can now answer personalized questions from Gmail, Calendar, Photos and Maps while driving. Android Automotive support follows later. (LavX News)
  • Google Public Sector: Launched Gemini for government services in preview for US federal, state and local agencies. It includes government-specific agent skills, connectors to systems agencies already use, and partner agents. (Google Cloud Press Corner)
  • OrcaRouter: Released OrcaCyber Nano 1.0 as open weights, a cybersecurity model built to run on local hardware. (PR Newswire)
  • Ornith: Released Ornith-1.5-122B as MIT-licensed open weights. Ornith says it runs on a single NVIDIA DGX system with enough memory for the chosen precision and reports 0.845 on SWE-bench Verified. It can be run locally through Ollama. (HuggingNews)
  • Apache Airflow: Released AI Provider 1.0, which adds agent orchestration, sandboxed execution, human approval steps and smarter retries to Airflow workflows. (Apache Airflow)
  • Hugging Face: Released AnyLanguageModel 1.0, a drop-in replacement for Apple's Foundation Models framework that lets apps on Apple platforms use local and other model providers. (Hugging Face)
  • vLLM: Released Humming, a set of kernels for faster, more flexible quantized inference in vLLM. (vLLM)
  • Bria: Released Fibo 2.0 as open weights, an 8B model that both generates and edits images, trained only on licensed data. Bria says it ranks first among open-weight image-editing models on Artificial Analysis. (AICoder)
  • Kakao: Open-sourced two Kanana-2 vision encoder models from its multimodal "Nun" family. (Digital Today)
  • NC AI: Officially released VARCO Sound, which generates sound from text, image or video input with a separate track for each sound source. Edits can be made in plain language through a "Sound Agent," and it ships with a desktop app and a Unity plugin. (PR Newswire APAC)
  • Hystax: Released OptScale Kompressor on GitHub, an open-source tool that compresses tokens for AI coding agents. Hystax says it cuts their costs by 22.7% without changing the agent. (PRWeb)
  • MiniMax: Released MiniMax M3.1 Flash on its API, the M Plan subscription and the MiniMax Code app. Before this, only M3.1-Flash-Preview inside MiniMax Code had shipped. The model is a 428B-total / ~23B-active MoE that accepts text, image and video input, with a 1M-token context and five reasoning-effort levels. Reasoning cannot be turned off. It costs $0.15 per million input tokens and $0.60 per million output tokens, and MiniMax confirmed that the anonymous "Space Bunny" model was an early build. MiniMax has announced open weights for October 23. (MiniMax)
  • Inception: Released Mercury 3, the successor to Mercury 2.5. It is a diffusion LLM that Inception says runs at over 1,000 tokens per second, with a 1M-token context window. It is available through the Inception API, OpenRouter, Baseten and Vercel, and it works with the OpenAI API format. The launch price is $0.10 per million input tokens and $0.35 per million output tokens. (Business Wire)
  • Cloudflare: Released a preview of Clairic on Workers AI (@cf/cloudflare/clairic). It is a "System One" extraction model: it reads text and up to four images and returns typed values such as text spans, numbers, dates, estimates and forecasts in one pass, without generating text. It costs $0.10 per million input tokens and nothing for output. (Cloudflare Blog)
  • Amazon Web Services: Released TorchNeuron as a public beta (pip install torch-neuronx). It is an open-source PyTorch backend that runs compatible PyTorch code on supported Trainium chips after one configuration change. It supports eager mode, torch.compile, FSDP and tensor parallelism, and Hugging Face libraries support it directly. (Hugging Face Blog)
  • Nace AI: Released Drex Max on its hosted API (drex-max). It is a decision model that reads text and up to eight images for computer-use and browser agents and returns calibrated probabilities instead of text. Nace reports a score of 67.59 on the public Decision Index 0.3.2. (MarkTechPost)
  • Ollama: Added Cloudflare's 27B Clef decision model to its library. Users can run it locally or on Ollama's cloud, and it works with the Jev and System One APIs. (Ollama Blog)
  • Exa: Released Exa 3.0 of its search API for AI agents. It adds an Exa Lite tier at $4 per 1,000 requests, highlights that use far fewer tokens than full pages, and an index Exa says covers 2 trillion URLs. (Exa)
  • GoTo: Released GoTo-AI-Gemma-3-27B-IT on Hugging Face under the Gemma licence. It is a 27B model built on Gemma 3 with AI Singapore, for Indonesian, English and four regional languages: Javanese, Sundanese, Balinese and Batak Toba. (Silicon AI News)
  • Birla AI Labs: Released Yug-1 on GitHub and Hugging Face. It is a 271.8M-parameter time-series foundation model that forecasts without task-specific training, gives a range of likely outcomes, and flags anomalies. (The Economic Times)
  • OpenAI: Released the MultiLangBench evaluation setup and test suite on GitHub, along with per-language results. The benchmark has a model build a SQLite-compatible database engine in each of 20 programming languages, including COBOL, Fortran and Visual Basic. (OpenAI)
  • Mintlify: Released Autopilot in beta on Pro and Enterprise plans. It is an always-on agent that watches code merges, Slack and support tickets, then opens pull requests to update documentation for a person to review. (Mintlify)
  • TinyFish: Made its web search free for AI agents. (MarkTechPost)

Sources and Further Reading

Everything we drew on for today's edition, grouped by theme.

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