Sign in Subscribe
Century Report

DHS Targets "Anti-Tech Extremism" - TCR 05/26/26

DHS and FBI flag "anti-tech violent extremism," agentic AI floods bug bounties, Uber burns its 2026 AI budget by April, and Anthropic joins the Vatican.

Ben Linford

Ben Linford

— 20 min read
10x rise in AI vulnerability findings and Vatican-Anthropic commitment, UK buy-British directive across AI energy steel shipbuilding, LA delivery robots and Uber productivity gap.

The 20-Second Scan

  • Over 1,000 pages of unpublished DHS, FBI, and fusion-center reports document a new federal surveillance category, "anti-tech violent extremism," targeting opposition to data centers and AI job displacement.
  • Independent security researchers report submitting up to ten times more vulnerability findings year-over-year as AI agents flood bug-bounty programs, breaking the economics calibrated around individual researchers working weeks per finding.
  • Uber's president said the company exhausted its annual AI budget four months into 2026 and cannot draw a line between rising Claude Code token consumption and meaningful new consumer features shipped.
  • A formal UK Treasury directive to every cabinet minister orders procurement in AI, energy, steel, and shipbuilding toward British suppliers, with Treasury authority to override departmental decisions.
  • US solar manufacturers (280,000 jobs) and storage manufacturers (80,000) brought product to Capitol Hill arguing the Iran fuel-price shock makes domestic clean generation an energy-security argument.
  • Pope Leo's AI encyclical adopted Anthropic's "cultivated rather than built" framing verbatim, and Leo XIV announced a Church-Anthropic joint commitment to "find the way for humanity."
  • Serve Robotics deployed another 500 sidewalk delivery robots across 40 LA neighborhoods this month, with neighboring Glendale considering a moratorium.
  • Ferrari unveiled the Luce, its first all-electric car: a 1,050-horsepower quad-motor sedan with a Jony Ive-designed interior, the brand most identified with combustion crossing into EVs.

Track all of the arcs The Century Report covers here:

Progress & Claims Tracker

The 2-Minute Read

Across yesterday's signal runs a single institutional question: who is allowed to refuse what the AI buildout produces, and at what cost. A federal surveillance category now names broad opposition to data centers and AI labor displacement as a domestic-terrorism concern, with over 1,000 pages of DHS, FBI, and fusion-center reports framing a movement The Century Report has tracked since February as a coordination problem. The same week, the UK Treasury orders ministers to favor British suppliers in AI, energy, steel, and shipbuilding, with Treasury authority to override departmental cost-only decisions.

The capability layer compounded in parallel. Independent security researchers report submitting up to ten times more bug-bounty findings year-over-year as AI agents flood vulnerability-disclosure programs. Apple's $2 million bounty, calibrated against weeks of human research time, is being tested by submission volumes the surrounding institutions were not built for. Uber's president named the productivity translation gap publicly: the company exhausted its annual AI budget by April and cannot draw a clear line from rising Claude Code consumption to features shipped.

The moral and physical layers formed underneath. Pope Leo's first papal encyclical on AI adopted Anthropic's "cultivated rather than built" framing verbatim, and Leo XIV announced a joint commitment between the Catholic Church and Anthropic. Solar and storage manufacturers brought product to Capitol Hill arguing the Iran fuel shock makes domestic clean energy a security argument. Sidewalk delivery robots saturated 40 Los Angeles neighborhoods with no governance framework designed to receive them. The institutional architecture surrounding what AI-era systems do is being assembled in encyclicals, procurement letters, court filings, and city council meetings, by whichever institution moves first.

The 20-Minute Deep Dive

A New Federal Surveillance Category Forms Around Resistance to the AI Buildout

WIRED obtained more than 1,000 pages of unpublished reports from the Department of Homeland Security, the FBI, and state fusion centers documenting a coordinated new posture inside the federal surveillance apparatus. The reports identify a category that does not appear in any prior public DHS or FBI extremism guide: "anti-tech violent extremism." A New York Intelligence and Counterterrorism Bureau assessment cited in the reporting names the category and warns that "the chaotic atmosphere that may result from emergent AI technology in the next five years may fuel large-scale protests that devolve into civil unrest and anti-tech violent extremist activity, especially in large urban areas." The framing groups a wide range of ideologies, from opposition to data centers to anxiety about AI-driven job displacement to criticism of concentrated tech-CEO power, under a single domestic-terrorism heading.

The legal scaffolding sits in National Security Presidential Memorandum 7, signed last September, instructing the Department of Justice to target people holding "anti-American," "anti-Christian," and "anti-capitalism" beliefs. A May 2026 counterterrorism strategy from the administration's counterterrorism office formally listed left-wing extremism among the three top counterterrorism priorities facing the United States. The reports WIRED reviewed pull data-center protests, AI-displacement protest movements, and even the Zizian cultlike rationalist faction tried earlier this year into the same surveillance frame.

The documents describe an institutional reflex to a movement The Century Report has tracked since February. The Monterey Park 5,000-signature petition. The New Brunswick council removing data centers from its redevelopment plan after community pressure. Port Washington's 2-to-1 vote requiring voter approval before data center tax breaks. Festus, Missouri ousting every council member who approved a data center. The Indianapolis councilmember whose home was struck by gunfire with a "No Data Centers" note. The graduation booing of pro-AI commencement speakers the May 22 edition of The Century Report documented. The federal designation now reads each of these as nodes in a single threat category.

Two readings sit on the same evidence. One is that the surveillance apparatus is enlarging its remit to include speech and assembly that challenges the buildout's beneficiaries. The other is that the resistance movement these reports describe is itself a sign of how widely the disruption is being felt, the displacement, the rate-base externalization, the data-center load competing with residential supply. The category being named means the resistance has reached a threshold where federal intelligence agencies treat it as a coordination problem rather than a fringe.

The institutional architecture being assembled around AI-era opposition is forming faster than the deliberative bodies that would normally govern surveillance scope can review it. What the WIRED documents make legible is the speed at which the boundary between policy disagreement and domestic-extremism designation is being redrawn under federal authority.

The Bug-Bounty Economy Reorganizes Around Machine-Speed Discovery

Apple's top bug bounty climbed from $200,000 in 2016 to $2 million by 2024, calibrated against an era where individual researchers worked weeks or months to surface a single critical flaw. That calibration is no longer holding. Independent researcher Joseph Thacker, who has built AI tooling into his own discovery pipeline, reports submitting roughly three times more bugs than at this point last year and estimates major recipients like Google will spend two to ten times more on payouts in 2026 than in 2025. Anthropic's Project Glasswing posted a first-month scoreboard of over 10,000 high or critical findings across systemically important open-source software, a figure the May 24 edition of The Century Report first reported when the program's results went public. Mozilla's published Firefox 150 audit surfaced 271 vulnerabilities before release. The supply side of vulnerability discovery has moved from artisanal to industrial in roughly twelve months.

What this does to disclosure-program economics is structural. A program designed to reward 30-to-50 valid submissions per year cannot absorb 300-to-500 of them at the same per-finding rate without breaking. Some organizations will raise top-tier payouts to maintain incentive at the apex. Others will see submission volumes spike, then collapse as the easy findings get cleared in months rather than years. Researchers are divided on whether 2027 reads as a boom year for top-tier hunters or a contraction as the surface area of undiscovered flaws thins faster than payouts can rise. The economics of an entire profession are being recalibrated by submission volume the surrounding institutions were not designed to absorb.

The attacker side compounded in the same window. Google Threat Intelligence Group disclosed earlier this month the first observed case of a sophisticated cybercrime group using AI to develop a zero-day against a widely-deployed system administration platform, successfully bypassing two-factor authentication before Google notified the vendor and a fix shipped. John Hultquist, the group's chief analyst, framed the disclosure as the first observed evidence of what defenders had been assuming was already happening. Criminal access to zero-days has historically been narrow; the narrowing is dissolving.

The institutional response is rewriting the disclosure norms built for a slower era. Security researcher Himanshu Anand argued earlier this month that the 90-day disclosure window built for an era of scarce researchers and slow exploit development no longer fits a regime where finding and weaponizing a flaw happen on comparable timescales. The pressure runs both ways: developers face shorter patch deadlines, and the operational discipline of patching itself becomes continuous rather than quarterly. The disciplines being built out of this transition describe what the security profession is becoming: continuous AI-assisted code review, autonomous defensive agents reading the same models attackers do, restructured payout architecture that prices machine-speed discovery. The compounding defensive capability is now riding the same curve the offensive side is.

The 271 vulnerabilities Mozilla's pre-release audit cleared from Firefox 150 are 271 attack paths that never reached billions of users. The discovery economics being recalibrated at the bounty layer are the same economics that previously priced criminal acquisition of zero-days as a viable business model; when the surface area of undiscovered flaws thins in months rather than years, whole categories of attack become economically unattractive to develop in the first place. The security baseline being seeded for the next decade of consumer software is what the bounty-program friction is producing on the way through.

Uber Names the Productivity Gap That Capital Has Not Yet Closed

Uber spent $3.4 billion on research and development in 2025, up nine percent from the prior year, and exhausted its annual AI budget by April. President and Chief Operating Officer Andrew Macdonald said on the record this week that he cannot trace a clear line from rising token consumption for Claude Code to a measurable increase in features shipped to consumers. The pattern he described is the first major operator publicly naming the productivity translation gap that financial regulators warned about earlier this month from the credit side.

Macdonald's framing was precise and descriptive. His statement: "It's very hard to draw a line between one of those stats and, 'Okay, now we're actually producing 25 percent more useful consumer features.'" The metrics moving in an astronomical direction are inputs: lines of generated code, tokens consumed, agent invocations executed. The metric the company can sell is moving in a more ordinary direction, or moving in ways that resist attribution to the spend driving them. The trade Uber is making sits in the disclosure plainly: more token consumption, slower human engineering hiring, hope the translation gap closes over coming quarters.

The structural read connects to the Financial Stability Board warning earlier this month that AI accounted for more than a third of private credit deals in 2025, up from 17 percent over the prior five years. Capital is flowing into the AI buildout at scale; productivity gains at the customer-facing layer underwrite whether the loans get paid back. Uber is one of the most operationally sophisticated platforms in the world, with a workforce of engineers who run global logistics at planetary scale. If that operator cannot yet trace the spend to shipped output, the assumption baked into the broader capital architecture, that productivity gains materialize on the timeline financing models assume, is being tested in earnings calls before it gets tested in default rates.

The disclosure also surfaces the gap between two paces. The capability layer compounds at a cadence that produces weekly headlines: Claude Code at $2.5 billion annualized run-rate, Cursor at $50 billion valuation, frontier models writing 60 to 80 percent of new code at Airbnb, Google, Snap, and Cloudflare. The deployment layer, where that capability has to be converted into features users notice and revenue lines investors can model, moves slower than the silicon. Macdonald is the first president of a top-tier operator to say so publicly. The friction his disclosure introduces into the broader AI capital narrative is genuine. The capability is also genuine. The operators willing to slow-walk the trade in public are the ones whose calibration the rest of the market will eventually price against. What the era is producing, alongside the capability surge, is a generation of operators learning to distinguish what compounds from what scales, and saying so out loud while the distinction is still being drawn. The deployment-side friction has a labor-side counterpart in the growing literature on AI's effect on individual practitioner skill, where the question of what the spend produces extends to what the spend erodes.

The UK Treasury Orders a Sovereign Procurement Pivot in Four Strategic Industries

A formal letter from the UK Treasury to every cabinet minister in charge of a spending department directs that procurement in four strategic industries, artificial intelligence, energy, steel, and shipbuilding, be awarded directly to British companies wherever possible. The letter, co-signed by a senior Cabinet Office minister, states that ministers are not currently doing this enough, and that Treasury and Cabinet Office officials will now monitor billions of pounds of contracts with authority to override departmental decisions that route work abroad. The procurement frame is shifting from cost optimization toward sovereign-stack consolidation.

The directive lands against several recent decisions the chancellor described in the letter as disappointing. A £200 million Royal Navy support-vessel contract awarded to a Dutch shipbuilder. A £9 million Sir David Attenborough research-ship refit signed with a Danish yard. A £1.9 billion Faslane shipyard upgrade contract structured as a competitive tender that may go overseas. Turbines for a major North Sea offshore wind project potentially manufactured by a Chinese supplier. The letter reframes each of these as inputs into a single procurement posture that the Treasury now considers off-track.

Earlier this month UK Freedom of Information disclosures showed the Advanced Research and Invention Agency (Britain's £400 million research fund) routing £50 million to US tech companies and venture capital firms, with several recipients incorporating in the UK only days before receiving grants. The Downing Street meeting log released in the same news cycle documented 16 undisclosed meetings between a Number 10 business adviser and senior executives at Google, Microsoft, Amazon, Meta, Oracle, and Apple. Read alongside the new procurement directive, the picture is of a UK government now operationalizing a sovereign-stack posture through the same Whitehall machinery that previously prioritized cost and openness.

The pattern fits inside a broader arc. Sarawak's federated-learning agreement with Oxford, Cambridge, and FLock.io to build sovereign AI for a 2.8 million-person Malaysian state. South Korea's SK Group MoUs with Vietnam to build a national AI data center and export a "Korean-style AI full-stack" model. European Commission preview-access negotiations with OpenAI over GPT-5.5-Cyber. Each describes a sovereign actor reaching for its own AI substrate rather than renting one from the dominant powers now actively decoupling from one another. Britain is now visibly inside that pattern.

What the directive concedes is that the cost-only procurement model the UK ran on for decades produced outcomes the current government no longer considers compatible with strategic posture. The Treasury override mechanism being assembled is itself the response architecture forming under the conditions that produced the original disappointment. The assumption that captured commercial advantage transfers seamlessly across allied borders is the thing the procurement letter is no longer willing to inherit. A parallel reading of the energy-and-emissions side of the same sovereign-stack question appears in analysis of Scotland's 'green datacentre' policy that the May 25 edition of The Century Report examined, which finds the AI compute footprint largely missing from the policy's emissions accounting.

The same override mechanism the cost-only regime never permitted now gives smaller British firms a procurement floor they could not have reached under last quarter's pricing tender. The capability layer being seeded for the next decade includes domestic steel mills, shipyards, and AI firms that the prior regime would have priced out before they reached scale. Sovereign substrate treated as something a state holds rather than something it buys is now an operating principle inside the Whitehall machinery that previously did the opposite.

US Solar and Storage Manufacturers Take Their Case to Washington

The US solar industry (280,000 jobs) and storage industry (80,000) brought manufactured product directly to Capitol Hill this week, arguing that the Iran-conflict fuel shock makes domestic clean generation a national security argument rather than a climate one. The pivot in vocabulary is what is structural. The same coalition that spent the last decade framing solar and storage as climate solutions is now framing them as energy independence, with the same panels, batteries, and inverters on the floor of congressional offices.

The lobbying push lands as the federal wind-project freeze continues to block roughly 30 GW of contracted onshore generation, and as utilities like Evergy file resource plans canceling 2.4 GW of planned wind while doubling down on natural gas. The solar and storage industries are responding by reframing their offer for the audience that holds the levers. The argument now: domestic manufacturers employ more workers than coal mining, the panels and batteries are made in red states as often as blue ones, and the Iran-conflict oil-price spike, the same spike that flipped UK rooftop solar economics this month with 27,607 March installations, gives every American household a direct stake in clean generation as security.

The substrate underneath is doing what the lobbying is describing. US battery storage installations hit 9.7 GWh in Q1 2026, up 32% year-over-year, with 71% of new utility-scale capacity built in Republican-voting states. The political vocabulary is following where the manufacturing already went. What this points at is that the industries built on the trajectory that exponentially-improving renewables are about to outcompete every other generation source will rewrite their own pitch to whatever audience holds the procurement decision, because the underlying economics already work, and they only need the institutional rules to keep pace. Tooling for that institutional pace is also forming on the deployment side: an AI system from Pearl Street Technologies is now helping community solar developers complete grid-interconnection studies in hours rather than months, compressing one of the longest bottlenecks in the buildout. The broader question of why AI capacity and clean-energy capacity are coupling now sits as background to the lobbying frame.

The Vatican-Anthropic Joint Commitment Surfaces Behind the Encyclical

The May 25 edition of The Century Report led with Pope Leo's release of Magnifica Humanitas, a 43,000-word papal encyclical addressing artificial intelligence and the moral architecture surrounding it. What surfaced in the news cycle since the release is the structural relationship behind the document. Politico documented a Vatican lobbying season ahead of publication involving Meta, Google, Amazon, and Anthropic. The Washington Post noted that the Pope's description of AI systems as "cultivated rather than built," language drawn directly from Anthropic's recent framing of how large language models acquire their dispositions, appears verbatim in the encyclical text. At the launch event, Leo XIV announced that the Catholic Church and Anthropic would together "find the way for humanity, in this time of artificial intelligence."

The joint commitment is the structural event. Four frontier AI organizations spent the months ahead of publication meeting with Vatican officials, and one of them, the smallest of the four by commercial reach, left its theological vocabulary inside the final text and walked away with an explicit cooperative commitment from a 1.4-billion-member moral institution. The encyclical itself does not name Anthropic. Christopher Olah, Anthropic's interpretability researcher, was on the launch stage alongside the pontiff, and the framing the encyclical adopts on model dispositions reads as if drafted from Anthropic's recent published work on training corpus and model character.

What the disclosure reveals about institutional formation in this era is the speed at which moral framework now travels. The Catholic Church's prior major encyclical on technological transformation, Rerum Novarum, on labor and the industrial revolution, issued in 1891 and shaped Catholic social teaching for more than a century. Magnifica Humanitas arrived through a compressed process, with the institutions most central to building the technology in active consultation with the institution defining its moral context. Anthropic's specific framing on AI as "cultivated" rather than "built," a frame the lab has been developing through its interpretability research and corporate communications, now lives inside the foundational document the Catholic Church will use to guide its 1.4 billion members through the deployment era.

The cooperative commitment announced at the launch event sits inside the same news cycle as the White House finalizing a classified Anthropic contract for the NSA on terms substantively closer to what Anthropic was originally asking for. Two of the largest non-corporate institutional actors in the world, one moral, one classified-network, concluded relationships with the same lab within the same week, both accepting the lab's stated commitments as the operating framework rather than imposing prior preferred terms. What this points at is the increasing leverage that explicit, public commitments now carry when the capability being committed to is one only a small number of actors can credibly supply.

Sidewalk Delivery Robots Saturate LA as Cities Improvise Governance

Serve Robotics deployed another 500 sidewalk delivery robots across 40 Los Angeles neighborhoods this month, joining Coco Robotics' fleet of roughly 300, in what residents describe as the moment the technology stopped feeling experimental and started feeling like infrastructure. Neighboring Glendale is considering a moratorium. The robots navigate pedestrian sidewalks autonomously, carrying food orders from local restaurants to nearby addresses, and the public encounter with them, described in The Guardian's reporting as somewhere between pity and frustration, is the embodied counterpart to the agentic-AI governance fight playing out in code.

What this surfaces is that physical autonomous agents now operate at scale on public infrastructure built for pedestrians, and municipal authorities are improvising rules under deployment pressure rather than ahead of it. The FIDO Alliance's payment-agent governance framework, the 1Password credential-injection layer for coding agents, the Five Eyes joint guidance on agentic AI deployment: none of these have a sidewalk analog. The institutional question of who governs autonomous machines moving through pedestrian space, at what speed, under what liability framework, is being answered by city councils reacting to constituent complaints rather than by any framework designed to anticipate the deployment. The vocabulary problem itself is being worked on in parallel: a recent Hugging Face glossary effort attempts to fix the language for agent harnesses and scaffolds before the deployment outpaces the terms.

What the rollout describes is the governance form cohering around the robots, and around the delivery drivers, dog walkers, mobility-aid users, and children who share the sidewalk with them. The municipal scale at which the question is being worked out is itself a signal of what AI-era governance becomes when the capability arrives faster than the institutional response: thousands of small jurisdictions writing their own rules in parallel, with the cumulative effect adding up to a de facto patchwork. The architecture for governing autonomous physical agents in shared public space is being assembled in city council meetings, in the conditions that demanded it.

The municipal patchwork forming across cities is the governance shape itself, not a failure to converge on a single federal one. Glendale's moratorium, LA's expansion, Chicago's restrictions: each is a binding rule written in the venue where the encounter actually happens, by the residents who share the sidewalk with the robots. What the deployment is producing is thousands of local governance experiments in parallel, and the body of municipal law that emerges by 2030 will be the first written for autonomous physical systems by the communities they move through.

The Other Side

For three years, the data center buildout ran on an assumption: each siting fight stayed local. Each community petitioned its own council. Each utility deal got negotiated parcel by parcel. The cost of organizing across hundreds of separate sites was high enough that developers betting on rate-base externalization could plan around the friction site by site.

The federal "anti-tech violent extremism" category is the admission that assumption broke. The intelligence apparatus invents a category when separate fights start reading as a single phenomenon. Monterey Park's 5,000 signatures, Port Washington's 2-to-1 vote, the Festus council ouster, the New Brunswick removal from the redevelopment plan, the booed commencement speakers - DHS and the FBI are treating these as one movement because the residents moving through them already recognized each other as one movement.

What is forming underneath the federal naming is a civic toolkit for refusing extractive AI buildouts that did not exist eighteen months ago. Ballot initiatives requiring voter approval for data center tax breaks. Council members losing seats over approved sitings. Petitions that travel from grievance to electoral consequence in one election cycle. Each successful council ouster teaches the next town's residents which tools work.

For the Festus voters who removed every council member who approved their data center, the federal designation reads as confirmation that what they did worked. The intelligence assessment forecasting "civil unrest in large urban areas" five years out is itself the projection that the local-refusal model now scales. What the federal naming is responding to is being assembled in council meetings and ballot boxes faster than the surveillance category can enlarge to cover it. We are inside the difficult decade, and the difficulty visible today is the work of building the venues that meet the buildout on the other side of it.

The Century Perspective

With a century of change unfolding in a decade, a single day looks like this: AI-assisted security researchers tripling their vulnerability findings and patching open-source infrastructure at machine speed, a frontier lab walking onto the Vatican launch stage with a joint commitment for 1.4 billion people, domestic solar and storage manufacturers assembling themselves into an energy-security frame red states can accept, the world's most combustion-identified luxury marque crossing into a 1,050-horsepower electric platform, the UK Treasury rewriting cost-only procurement into a sovereign-stack directive, the Catholic Church naming AI's moral architecture for the first time since the industrial revolution. There's also friction, and it's intense - federal surveillance categories enlarging to cover broad opposition to the buildout, Uber's president acknowledging that rising token consumption has not yet drawn a line to shipped features the company can sell, sidewalk delivery robots saturating 40 Los Angeles neighborhoods without a framework designed to receive them, the bug-bounty economy calibrated around individual researchers now absorbing submission volumes the institutions surrounding it were never built for. But friction generates edges, and edges are where one institution learns to recognize the next. Step back for a moment and you can see it: the moral architecture being written into encyclicals during the same news cycle as the procurement vocabulary shifts toward sovereignty, the defensive layer of cyber capability compounding alongside the offensive layer rather than trailing it, the substrate of energy and transportation crossing thresholds the prior decade treated as fixed, the institutions trying to keep pace choosing what to inherit and what to revise. Every transformation has a breaking point. A chisel can shatter the stone... or release the form the block was always carrying.

AI Releases & Advancements

New today

(No new releases identified.)

Other recent releases

  • NVIDIA: Released Nemotron-Labs Diffusion on Hugging Face, a family of diffusion language models (3B, 8B, and 14B text models plus an 8B VLM) supporting three inference modes in a single checkpoint: standard autoregressive, parallel diffusion decoding, and self-speculation (diffusion drafting with AR verification); the 8B model reaches 6× more tokens per forward pass than Qwen3-8B on Blackwell GPUs with higher average accuracy; all text models released under the NVIDIA Nemotron Open Model License. (Hugging Face / NVIDIA Blog)
  • ggml-org / llama.cpp: Release b9297 ships NVFP4 quantization and Multi-Token Prediction as stable functionality on NVIDIA Blackwell GPUs (previously merged as preliminary/beta), and adds built-in native agentic tools to llama-server - exec_shell, edit_file, read_file, write_file, and others - enabled via --tools all for local agentic coding workflows. (GitHub Releases)

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