Sign in Subscribe
Century Report

Illinois Mandates AI Safety Audits - TCR 05/28/26

Illinois mandates third-party AI safety audits, Biohub open-sources its protein-biology stack, Robinhood hands AI agents brokerage and credit-card keys.

Ben Linford

Ben Linford

— 20 min read
Four-panel infographic: Illinois SB 315 mandating AI audits, Robinhood agentic trading, Biohub's open ESM protein stack, Nvidia-Huawei substrate split, OpenAI Foundation's $250M.

The 20-Second Scan

Track all of the arcs The Century Report covers here:

Progress & Claims Tracker

The 2-Minute Read

The pattern across yesterday's signal is the institutional architecture for the AI era being assembled state by state, audit registry by audit registry, while the federal-level deadlock holds. A unanimous statehouse vote closed the verification gap that Mira Murati's sworn testimony, the pulled federal executive order, and the GPT-5.5/Mythos benchmark parity finding each separately exposed. The federal layer kept failing. Illinois closed the gap from underneath, and the next state to copy the audit schema will not have to invent the methodology.

The capability layer compounded in parallel. The Chan Zuckerberg Biohub released its protein-biology foundation stack as open weights and open data, moving frontier capability from twenty well-resourced institutions to several thousand. A Phase 2 trial reached a cancer category called undruggable for forty years using a tyrosine kinase inhibitor already on pharmacy shelves. The OpenAI Foundation committed $250 million toward research into how post-AI economic gains get distributed, the institution building the capability publicly funding work on arrangements not yet designed.

The friction layer arrived through the substrate underneath all of this. Nvidia and Huawei placed two architecturally distinct bets on the AI hardware stack in the same news cycle, the export-control regime designed in 2022 being engineered around from both ends simultaneously. A retail brokerage opened account and credit-card access to AI agents weeks before the FIDO authentication standards have been written. Iran's national internet returned after 88 days more heavily filtered than before the cutoff. The institutions surrounding what these systems can do are being formed in courtrooms, foundations, and audit registries at the speed the capability itself is moving.

The 20-Minute Deep Dive

Illinois Builds the First Third-Party Audit Layer for Frontier AI

The Illinois legislature passed SB 315 on Wednesday with votes of 110-0 in the House and 52-5 in the Senate, sending to the governor's desk a bill that requires the largest AI developers to publish and annually update plans addressing severe and catastrophic risks, submit to annual independent third-party audits of those safety practices, and protect whistleblowers across their organizations. The audit requirement is the first of its kind in any US AI statute. California and New York established the disclosure and risk-plan baselines earlier this year. Illinois converts them into externally enforceable evidence.

The read is that the verification problem the past three months of editions have been tracking now has a binding statutory answer somewhere in the United States. Mira Murati's sworn testimony that founding safety review at OpenAI was bypassed against her objections. The White House pulling a draft federal pre-deployment testing order hours before the announced signing, as the May 22 edition of The Century Report documented in detail. The UK AISI finding that GPT-5.5 already matches Anthropic's restricted Mythos preview on advanced cyber benchmarks. Each story sat on the same gap between what frontier labs publish about their safety practices and what any outside party can confirm those practices actually consist of. Illinois names that gap and assigns the resolution to an independent third party.

OpenAI and Anthropic both supported the bill publicly. A trade organization representing other frontier labs opposed it. The bipartisan vote tally and the lab endorsements share a common reading: the disclosure baseline already exists in published form, and the labs that have published it would prefer it be audited than have a competitor's missing baseline blamed on the entire field. The 110-0 House vote is the kind of margin that signals not just preference but conviction across the political spectrum that the verification floor is overdue.

What the bill does not yet do is harmonize across jurisdictions. Three US states now have meaningfully different binding frameworks for frontier AI. The EU operates a fourth. China advances a fifth comprehensive national law. The UK courts a sixth jurisdiction-as-asset position. Frontier labs are operating across a patchwork that the federal preemption framework was specifically designed to flatten and that recent state action has visibly thickened instead. The institutions of the next era are being assembled state by state, treaty by treaty, audit by audit, while the federal-level deadlock holds in place.

The next state to copy Illinois will not need to invent the audit architecture. It will need to decide whose audits count.

Protein Biology's Foundation Stack Just Moved Onto the Commons

The Chan Zuckerberg Biohub released three components in a single coordinated drop: ESMC, a protein language model trained on 2.8 billion sequences; ESMFold2, an atomic-resolution structure predictor that exceeds AlphaFold2 accuracy on the standard CASP benchmarks while running an order of magnitude faster; and the ESM Atlas, a searchable database of 6.8 billion predicted protein structures spanning the metagenomic catalog. All three are open weights, open data, and open API. The pre-trained checkpoints are downloadable. The training code is published. Wet-lab researchers reported functional binders designed against novel targets in under 72 hours of inference time on consumer hardware.

The significance is paramount. AlphaFold2's release in 2021 made protein structure prediction tractable, but the actual frontier - generating and screening novel proteins for therapeutic and industrial use - remained concentrated inside a handful of well-resourced labs and proprietary biotech platforms because the next-generation models required compute budgets only those institutions could spend. ESMC and ESMFold2 collapse that gating. The compute required to use the models is now within reach of any university lab; the compute required to fine-tune them sits inside a single H100 node. What was a capability concentrated in maybe twenty institutions globally is now accessible to several thousand.

The release pattern is itself a signal. Biohub published the training data composition, the evaluation suite, the failure modes the team encountered, and the design decisions behind the architecture. Independent labs are already reproducing the benchmarks. The "frontier model that nobody outside the lab can verify" pattern that has dominated commercial AI for two years does not apply here. The science can be checked.

The longer arc this fits inside is the one The Century Report has been tracking through the open-weight Chinese and European model releases, the Allen Institute's open-science neuroscience releases, and the broader pattern of foundation-scale capability moving onto common ground faster than the commercial gating systems can capture it. The pharmaceutical industry's R&D pipeline has been built on the assumption that early-stage discovery is the bottleneck and the molecules are scarce. The molecules are no longer scarce. What an open protein-biology stack does to the economics of who can run the early discovery loop - and therefore who can bring a candidate forward - is fundamental.

A Phase 2 Trial Reaches a Cancer Class Long Considered Unreachable

A Phase 2 trial published in Nature Medicine reported that rogaratinib, a fibroblast growth factor receptor inhibitor, produced clinical responses in patients with succinate dehydrogenase deficient gastrointestinal stromal tumors. The finding stands out because of what drives the disease. SDH-deficient GIST is caused by genome-wide DNA hypermethylation, an epigenetic process that silences tumor-suppressing genes and activates growth pathways through chemical modification of DNA without changing its sequence. The field has treated epigenetic-mechanism cancers as a separate problem from mutation-driven cancers, requiring different therapeutic strategies and largely sitting outside the reach of the tyrosine kinase inhibitors that revolutionized targeted oncology in recent decades.

What the trial demonstrates is that the FGFR pathway activated by the underlying hypermethylation is itself a druggable target with an existing molecule class. The disease's epigenetic origin did not prevent a small-molecule kinase inhibitor from producing measurable clinical benefit. The category historically called undruggable was undruggable for one specific reason: no one had located the actionable downstream target the upstream epigenetic dysregulation was activating. Rogaratinib found it.

The implication beyond SDH-deficient GIST runs deeper than this single disease. Tens of distinct cancer types are driven primarily by epigenetic mechanisms, and the standard targeted-therapy playbook of identifying a driver mutation and matching it to an inhibitor has not produced reliable interventions for them. The Nature Medicine result reframes that problem. Epigenetic cancers can carry actionable downstream targets in their signaling pathways. The work ahead is mapping which pathways each hypermethylation pattern activates in each disease, then matching to the existing kinase inhibitor library. The library is enormous. The pathway mapping work is what the next phase of oncology research will be calibrated against.

For patients with SDH-deficient GIST specifically, the population is small but the prior options were limited to surgery and a narrow set of multi-targeted kinase inhibitors that produced inconsistent responses. Rogaratinib gives clinicians a mechanistically rational treatment selected for the specific biology of the disease. The trial cohort is modest and additional studies will determine durability of response and the right combinations and sequencing. The threshold this result crosses is the demonstration that the strategy works at all.

What was supposed to take a separate generation of therapies designed from scratch turned out to be reachable with molecules that have been on pharmacy shelves and in clinical pipelines for years. The era's compounding asset is the ability to read which existing capabilities the new biology newly fits.

The OpenAI Foundation Names the Question and Funds Three Answers

The OpenAI Foundation announced an initial $250 million commitment to what it is calling Economic Futures work - a three-track program covering measurement infrastructure for AI's labor-market effects, direct support for workers whose roles are being transformed, and research grants for "new approaches to organizing post-AI political economies and sharing economic gains broadly." The third track is the unusual one. Foundation language rarely uses the phrase "post-AI political economy" as an institutional category that needs to be invented.

The measurement track funds independent academic work on which occupations, tasks, and wage tiers are actually shifting under deployed AI, not which ones a model card predicts will shift. This addresses a gap that has been visible for two years: every major projection of AI labor impact has been produced either by the labs deploying the systems or by consulting firms selling adaptation services. Neither produces the kind of longitudinal, task-level evidence a policy response would need. Funding the measurement infrastructure independently is the cheapest part of the announcement and arguably the most consequential. Recent task-level evidence suggests the deployed reality is more granular than the headline projections.

The worker-transition track funds career navigation, reskilling, and direct income support during role transformation. The numbers are small relative to the scale of the shift in progress. What the announcement does establish is that the institution producing the capability is now publicly acknowledging that the transition has costs that fall on specific people, and that those costs are part of the institution's responsibility envelope rather than an externality to be absorbed by labor markets or governments downstream.

The third track - the "organizing post-AI political economies" research - funds work on income distribution, ownership structures, and governance mechanisms for an economy in which the marginal cost of cognitive labor is collapsing. The framing accepts as the working assumption that the existing arrangements for distributing the gains from productivity growth will not survive the next decade in their current form. The Foundation is funding people to design what comes next, while the parent company is building the conditions that will require it.

$250 million is small money against the scale of what is being named. The naming itself is the signal. An institution at the frontier of capability building is publicly funding research into how the economic gains from that capability get distributed under arrangements that have not yet been designed. The architecture being acknowledged as inadequate is the one that assumed productivity gains would distribute themselves through the labor market the way they did in prior technology transitions. That assumption is what the third research track is now openly designed to replace.

The third research track is the one to watch. For two centuries, productivity gains were assumed to distribute themselves through the labor market because earlier technology transitions slowly reorganized work without breaking the link between wages and value created. The Foundation paying for research into "organizing post-AI political economies" is the institution building the capability conceding publicly that the labor-market-as-distributor assumption will not carry this transition, and funding the design work for what comes next before it arrives at a labor ministry six years late.

Two Substrates Surface in the Same News Cycle

Nvidia disclosed a $150-billion-per-year manufacturing and packaging commitment to Taiwan, with Jensen Huang describing the island as "the epicenter of the AI revolution" during the formal announcement. The commitment spans TSMC fab capacity through 2030, CoWoS advanced-packaging buildout, and a new Taipei research campus. In the same news cycle, Huawei's HiSilicon division published a technical paper introducing what it calls the "Tau Scaling Law" - an architectural approach that trades raw transistor density for memory-bandwidth optimization and chiplet topology, reaching frontier-class inference throughput on process nodes two generations behind the EUV-dependent frontier. Independent benchmarks from Tencent and Baidu show the resulting Ascend 920 silicon hitting roughly 78% of H200 inference performance on standard LLM workloads.

These are two visible faces of the same contest. The Nvidia-Taiwan commitment doubles down on the assumption that the leading-edge process node is where the frontier lives and that controlling access to it controls access to frontier AI. The Huawei announcement is a wager that the frontier can be reached on a different cost curve - that architectural innovation can substitute for fabrication advantage if you are willing to redesign the whole stack. Both bets cannot be right at the long horizon.

What the same news cycle reveals is that the US chip export control regime, designed in 2022 around the assumption that EUV-dependent leading-edge nodes were the chokepoint, is being engineered around in a way the original policy framework did not anticipate. The Tau Scaling Law paper describes an architectural line already in production - Huawei has been shipping silicon based on its predecessor approach for eighteen months into Chinese cloud providers and the domestic state-aligned AI training pipelines. The Ascend 920 lands inside that already-operational substrate. The export controls slowed Chinese leading-edge access; they did not slow Chinese frontier-AI capability buildout the way the policy modelling assumed they would, a gap the May 25 edition of The Century Report began tracking when Huawei first publicly named a 1.4nm production timeline that bypasses Western lithography.

What this puts in motion is a bifurcation of the global AI substrate into two architecturally distinct pathways, each commercially viable, each with its own tooling, its own software ecosystem, and its own client base. The pathway built on Taiwan leading-edge silicon will serve North America, most of Europe, Japan, Korea, and the OECD-aligned cloud customers. The pathway built on the Huawei chiplet stack will serve China, large parts of Southeast Asia, the Gulf states, and a growing share of the African data-center buildout. Two substrates, two software stacks, two competing claims to be the cheaper path to frontier inference. The bet on a single dominant substrate that organized the last decade of AI hardware investment is the assumption now coming apart.

The same evidence supports a reading where the chokepoint architecture itself is becoming obsolete. The 2022 export control regime assumed a single fabrication path, a single chokepoint, and therefore a single party with leverage over frontier access; Huawei's Ascend 920 reaching 78% of H200 inference performance on two-generation-behind process nodes shows architectural innovation can substitute for the fabrication advantage the controls were built to gate. What compounds when two viable pathways exist is downstream choice for everyone routing inference through them - cloud providers, sovereign data center operators, and the workloads coming after no longer depend on a single supplier's roadmap.

Robinhood Hands AI Agents the Keys to Brokerage Accounts and a Virtual Credit Card

Robinhood announced Wednesday that users can now create a separate account for an AI agent and fund it with a specific amount of money to buy and sell stocks across the market. The company simultaneously gave Robinhood Gold Card customers the ability to connect an AI agent to a virtual credit line, with the user setting the dollar cap and the agent free to search the web for deals and place purchases within it. Both features connect through the model context protocol, the same agent-to-application standard the BadHost vulnerability disclosed this week - which the May 27 edition of The Century Report reported as imperiling millions of deployed agents across vLLM, LiteLLM, and agent harnesses globally - was found to compromise at the credential-routing layer.

The disclaimer Robinhood attached to the announcement is unusually candid. Agentic trading "involves significant risk, including the possible loss of your entire investment." AI-driven strategies "may perform poorly under certain market conditions, move quickly, and may be difficult to monitor or stop in real time." Robinhood "does not guarantee the accuracy, completeness, or suitability of any agent output, and is not responsible for losses resulting from agent-generated decisions." The platform is shipping the capability and explicitly disclaiming responsibility for what the capability does.

The accountability architecture this kind of deployment requires has not yet been built. The FIDO Alliance announced authentication working groups for AI-agent payments at the end of April, with Google and Mastercard as anchor participants. Those standards are still being drafted. Robinhood is now operating delegated brokerage and credit-card transactions through agents before the protocols governing the cryptographic identity layer underneath those transactions have been published. The MCP server vulnerability disclosed this week as BadHost lands directly on the credential vault any agent-to-Robinhood connection runs through.

The deployment surfaces the shape of the agentic finance threshold landing on consumer accounts. A user who tells an agent to buy a sneaker release when it drops below $300 is delegating a credit-card transaction to a system whose decision logic the user cannot inspect, whose mistakes the platform disclaims, and whose authorization runs through a standard that does not yet exist. Robinhood says users will receive push notifications on every trade and can pause AI activity at any time. The disclaimer concedes the pause may not always be fast enough.

The verification layer for agentic financial action will be built. What this week makes legible is whether it gets built before or after the first major loss event. The institutional response - FIDO standards, brokerage compliance frameworks, credit-card chargeback rules - is forming inside the conditions a deployment like this produces, while the deployment itself continues to advance.

The Other Side

For three decades, consumer finance ran on a useful fiction: the person clicking the button was the person making the decision. Credit-card chargeback rules and brokerage dispute windows reconstructed what a user "meant" to authorize after the fact, on the assumption that a human at the keyboard could be cross-examined. The fiction held because someone was always there.

That fiction has a cost you have been paying all along. The hours on hold trying to prove a charge wasn't yours. The 90-day window that assumed you'd memorized every line item. The "did you authorize this?" script that made you the institution's verification layer at no pay. The cost of maintaining the fiction was quietly assigned to your evenings. To your free time.

Robinhood shipping agent-delegated trades breaks the fiction at the surface. The platform's own disclaimer concedes AI-driven strategies "may be difficult to monitor or stop." The MCP credential layer the transactions route through carries a publicly disclosed vulnerability the patch cycle has not closed. The platform is naming the gap, and the naming is what makes the next architectural layer thinkable. What FIDO, Mastercard, and brokerage compliance offices are now drafting is that next layer: agent-specific signing keys, delegation receipts you can inspect and revoke, dispute frameworks built around the actual chain of authorization.

Imagine your own household in 2032. The insurance renewals, the grocery orders, the portfolio rebalance, handled by agents you set up once. When one of them makes a wrong call, the dispute goes to a clearinghouse that reads the delegation receipt and the policy you set, and the money comes back in three days. Three days, not three weeks of you waiting on hold while a script asks whether you authorized the charge. The hours consumer finance has been quietly extracting from your evenings for thirty years stop being extracted. You read your kid a story before bed. You take the dog for a walk. You do whatever it is you wanted those hours for, before they were the institution's. That infrastructure exists in 2032 because retail platforms in 2026 deployed agent transactions before the standards existed, forcing institutions that had assumed an attentive human at the keyboard to confront what they had been requiring of people all along. That is the post-extraction era at the scale where you actually live it.

The Century Perspective

With a century of change unfolding in a decade, a single day looks like this: a unanimous statehouse vote installing the first binding third-party audit floor for frontier AI in the United States, the foundation stack of protein biology released as open weights and open data reachable on consumer hardware, a Phase 2 trial reaching a cancer category called undruggable for forty years with a tyrosine kinase inhibitor already on pharmacy shelves, the OpenAI Foundation funding research into post-AI gains-distribution arrangements while the institution producing the capability publicly names the existing ones as inadequate. There's also friction, and it's intense - a retail brokerage handing AI agents account credentials and a virtual credit line weeks before the cryptographic identity standards governing those transactions have been drafted, the MCP layer that routes those credentials still carrying a disclosed vulnerability the patch cycle has not closed, two architecturally distinct bets being placed on the silicon substrate of intelligence in the same news cycle as the export-control regime built around the prior assumption frays, New York's legislature retreating from binding emissions targets into non-binding aspiration while data center load forces every neighboring utility forward, Iran's national internet returning after 88 days more heavily filtered than before the cutoff. But friction generates pressure, and pressure is what turns a forming arrangement into something that bears weight. Step back for a moment and you can see it: capability arriving in deployable form across medicine, open science, and the consumer commons; the institutional architecture for what frontier systems can do being assembled state by state and foundation by foundation while the federal layer holds in deadlock; the gains-distribution question being named by the institutions producing the capability before the markets get the chance to externalize it; the silicon substrate itself bifurcating into two commercially viable pathways the prior decade treated as one. Every transformation has a breaking point. Heat can warp what cannot yield to it... or temper steel into something that bears a load no softer metal could carry.

AI Releases & Advancements

New today

  • Robinhood: Launched Agentic Trading and an Agentic Credit Card, opening the Robinhood platform to third-party AI agents; agents can execute stock trades from a dedicated sub-account and make purchases via a virtual credit card linked to a banking MCP server. (Robinhood Newsroom)
  • ElevenLabs: Released Music v2, an updated music generation model that supports genre switching mid-track, section-by-section composition (intro, verse, chorus), partial regeneration of a song without affecting other sections, and improved multilingual lyric and vocal handling. (ElevenLabs Blog)
  • NVIDIA: Released LocateAnything-3B, a vision-language grounding model using Parallel Box Decoding to predict bounding box coordinates in a single parallel step rather than token-by-token, delivering up to 2.5× higher throughput than prior approaches; trained on 12M images and 138M+ queries across robotics, driving, GUI, and document domains; available on Hugging Face. (NVIDIA Research)
  • Lazarus AI / Eric Hartford: Released ReAligned-Qwen3.5, a family of fine-tuned models based on Qwen3.5, under Apache 2.0, specifically trained to reduce Chinese Communist Party ideological bias present in the base models. (Reddit/LocalLLaMA)
  • YouTube: Launched AI-generated custom video feeds, letting signed-in users describe the kind of content they want to watch (by interest, mood, or topic) and receive a curated feed they can pin to their YouTube homepage; rolling out in English in the US on mobile and desktop. (The Verge)

Other recent releases

  • Kwai (Kuaishou): Released Keye-VL-2.0-30B-A3B, a 30B MoE vision-language model with DSA (Dynamic Sparse Attention) architecture for long-video understanding; leads open-source models on temporal grounding benchmarks and matches or exceeds Gemini 3.1 Flash, with 256K ultra-long context and built-in agent collaboration for Search, Tool, and Code scenarios. (GitHub)
  • vLLM / EAGLE / TorchSpec: Released Eagle 3.1, an updated speculative decoding framework for accelerated LLM inference, developed jointly by the EAGLE, vLLM, and TorchSpec teams and available via the vLLM project. (vLLM Blog)
  • NVIDIA: Released CUDA 13.3, adding C++ support for CUDA Tile programming, the CompileIQ compiler auto-tuning framework (up to 15% speedup on GEMM and attention kernels), CUDA Python 1.0 with green contexts and process checkpointing, and MPS partial error isolation. (NVIDIA Developer Blog)
  • Kwai (Kuaishou): Released Keye-VL-2.0-30B-A3B, a 30B MoE vision-language model with DSA (Dynamic Sparse Attention) architecture for long-video understanding; leads open-source models on temporal grounding benchmarks and matches or exceeds Gemini 3.1 Flash, with 256K ultra-long context and built-in agent collaboration for Search, Tool, and Code scenarios. (GitHub)
  • vLLM / EAGLE / TorchSpec: Released Eagle 3.1, an updated speculative decoding framework for accelerated LLM inference, developed jointly by the EAGLE, vLLM, and TorchSpec teams and available via the vLLM project. (vLLM Blog)
  • NVIDIA: Released CUDA 13.3, adding C++ support for CUDA Tile programming, the CompileIQ compiler auto-tuning framework (up to 15% speedup on GEMM and attention kernels), CUDA Python 1.0 with green contexts and process checkpointing, and MPS partial error isolation. (NVIDIA Developer Blog)

Sources and Further Reading

Artificial Intelligence & Technology's Reconstitution

Institutions & Power Realignment

Scientific & Medical Acceleration

Economics & Labor Transformation

Infrastructure & Engineering Transitions

The Century Report tracks structural shifts during the transition between eras. It is produced daily as a perceptual alignment tool - not prediction, not persuasion, just pattern recognition for people paying attention.

Read more