OpenAI’s unverified plan, Seoul’s contradiction, rural America’s veto

TL;DR

• OpenAI’s five-pillar cyber plan ships with UK AISI unable to verify the GPT-5.5 jailbreak patch before launch.
• Korea’s MSIT cut Naver from its ₩530B Sovereign AI program, then signed an MOU embedding Google DeepMind’s proprietary stack.
• Samsung is locked in for over 60% of Google’s 2026 HBM under price floors and 10–30% prepayments.
• $162B in US data center projects sit blocked or delayed since 2023, with $98B frozen in Q2 2025 alone.
• Virginia recoups roughly 48 cents per dollar of foregone sales tax on data center hosting.

Today’s frontier news isn’t about a model launch or a megadeal — it’s about the people AI’s growth thesis quietly assumed would go along with it. OpenAI publishes a five-pillar cybersecurity plan that leans on independent validators, on the same day UK AISI red-teamers say the validation pipeline doesn’t actually work. South Korea’s MSIT, two weeks after cutting its national champion from a sovereign-AI program for lacking originality, signs an MOU importing Google DeepMind’s proprietary stack — and locks Samsung into majority HBM supply on Google’s terms. And in rural America, a cross-ideological coalition has now blocked or delayed $162B of hyperscaler buildout, enough that the freezes are showing up in earnings guidance.

Three very different stories, one frame: the constituencies AI took as fixed scenery — security evaluators, sovereign allies, communities that host the racks — are renegotiating the deal in public. The infrastructure side is the loudest, but the pattern is what to watch.

OpenAI’s cybersecurity action plan lands with its safety stack unverified

Source: openai-blog · published 2026-04-29

TL;DR

• OpenAI’s five-pillar “Intelligence Age” cyber plan leans on a defender’s-advantage thesis that independent researchers say isn’t supported.
• UK AISI red-teamers cracked GPT-5.5 in six hours; a misconfigured eval left them unable to verify the patch before launch.
• The plan’s FedRAMP Moderate badge sits on a program ProPublica calls a “rubber stamp,” and never mentions the Pentagon-contract backlash reshaping public trust.
• Read it as a positioning document, not a technical roadmap.

The pitch

OpenAI’s Action Plan frames AI as a structural win for defenders: faster vulnerability remediation, a ~40% drop in simulated dwell time via Codex-driven incident response, and better phishing filters via the GPT-5 family. Five pillars — democratized defense, government coordination, frontier-model security, deployment monitoring, and end-user tooling — are wrapped around a new FedRAMP Moderate authorization that opens the door to U.S. federal adoption.

That’s the story OpenAI is telling. The story the rest of the cybersecurity press is telling around it is considerably less flattering.

What the announcement omits

Transformer’s Shakeel Hashim puts the governance objection bluntly:

OpenAI shouldn’t be deciding if its GPT-5.5 is safe enough to release ⁵

That’s the load-bearing critique. The Frontier Capability Security pillar is, in practice, self-attestation — and the one external auditor with access publicly says it couldn’t finish the job ¹.

The defender’s-advantage thesis is contested

The “AI flips asymmetry toward defenders” framing is the report’s intellectual spine, and it’s the claim independent researchers most directly reject. Discovery has gotten cheap; deployment hasn’t. Patching enterprise systems is still months of human-led testing and rollout, which means lower-skilled attackers — the population AI uplifts most — sit inside an exploitation window that gets wider, not narrower, as discovery accelerates ³. The 40% dwell-time number from OpenAI’s controlled pilots doesn’t speak to that regime at all.

Anthropic’s Project Glasswing — gated, consortium-only access to a frontier defensive model — represents the competing bet: that “democratizing” frontier cyber capability to thousands of vetted defenders is the wrong threat model when the same API can be turned around. OpenAI’s plan doesn’t engage with that critique.

The trust backdrop the post ignores

The action plan arrives mid-crisis. The #QuitGPT campaign over OpenAI’s Pentagon contract drove roughly 2.5M uninstalls and the resignation of robotics lead Caitlin Kalinowski, who said “surveillance of Americans without judicial oversight and lethal autonomy without human authorization are lines that deserved more deliberation than they got” ⁶. The Persona/FinCEN allegations ⁴ mean the KYC layer underpinning “Trusted Access for Cyber” is itself the subject of a surveillance investigation.

None of that is acknowledged in the post. Read in isolation, it’s a confident roadmap. Read against the week’s reporting, it’s a positioning document — pitched to federal buyers — that asks readers to take the security claims on faith precisely when the only outside auditor with access has said it can’t.

What’s actually at stake

If FedRAMP Moderate becomes the de facto bar for frontier-model deployment in U.S. agencies, the AISI verification gap ¹ and the ProPublica structural critique ² stop being commentary and start being the regulatory regime. The action plan is the opening move in that fight.

---

DeepMind’s Korea deal cuts against Seoul’s “Sovereign AI” doctrine

Source: deepmind-blog · published 2026-04-27

TL;DR

• Google DeepMind signed an MOU with Korea’s MSIT to embed AlphaFold, AlphaGenome, AlphaEvolve and WeatherNext into national research.
• Two weeks earlier, MSIT cut Naver from its ₩530B “Sovereign AI” finals for lacking “originality” — now it’s importing Google’s proprietary stack.
• A parallel hardware deal locks Samsung in for >60% of Google’s 2026 HBM, with price floors and 10–30% prepayments.
• Hassabis pitched 5-year AGI at the Lee Sedol reunion; Lee warned against ceding “sovereignty over thought.”

The announcement, and the contradiction it papers over

Google DeepMind’s new partnership with Korea’s Ministry of Science and ICT puts AlphaFold, AlphaGenome, AlphaEvolve, AI Co-scientist and WeatherNext inside the national research pipeline, anchored by an “AI Campus” in Seoul and a National AI for Science Center opening May 2026. The blog frames Korea as a “global leader in AI innovation density.” It does not mention that two weeks earlier, MSIT eliminated Naver Cloud and NC AI from the ₩530 billion “Sovereign AI” finalist round, with expert panels citing a lack of technical “originality” in Naver’s foundation model ⁷.

The same ministry that disqualifies domestic champions for derivative work is now wiring a foreign proprietary stack into bio, materials, and grid-scale weather forecasting. The 85,000 Korean researchers already using AlphaFold is a real number — and exactly the dependency Korea’s sovereign-AI rhetoric was meant to prevent.

Follow the money — and the HBM

The K-Moonshot envelope DeepMind alludes to is larger than the post lets on: ₩10.1 trillion (~$7.27B) for AI in 2026, a 206% YoY jump, with ₩464B specifically earmarked between 2027 and 2031 to build independent foundation models in biotech and materials ⁸. The DeepMind partnership runs straight through that same domain.

It also runs alongside a much harder commercial integration. Samsung has reportedly locked in more than 60% of Google’s 2026 HBM allocation for Ironwood TPUs, under multi-year long-term agreements that include unprecedented price floors and 10–30% prepayment deposits from Google ⁹. The “AI Campus” is the soft-power face of a supply-chain marriage with Samsung and SK hynix that the science MOU helps cement politically.

Two playbooks for “national” AI

DeepMind’s National Partnerships for AI is now visibly diverging from OpenAI’s approach:

Korea got the science-hub version. Whether that’s compatible with funding ₩464B of “independent” foundation models in the same sectors DeepMind is now servicing is the open question MSIT hasn’t answered.

Dissent, on both sides of the Pacific

At the AlphaGo 10-year commemorative event, Demis Hassabis estimated AGI within five years. Lee Sedol, the symbolic counterweight, warned that humans must not cede “sovereignty over thought” or control of creative processes to machines ¹¹.

“Incredibly ashamed.”

That’s how some of the 600+ DeepMind staff described their own employer this year, after protesting a classified U.S. Department of Defense contract and the quiet removal of prior pledges against military applications ¹². The Korean AI Safety Institute is now a designated collaborator on “frontier AI risks” under this MOU — a partnership whose credibility depends on a safety culture DeepMind’s own researchers are publicly questioning.

The deal is simultaneously a scientific coup for Korean labs, a hardware-supply consolidation for Google, and a quiet repudiation of Seoul’s sovereign-AI line. The DeepMind blog post acknowledges only the first.

---

Rural America is now pricing into hyperscaler guidance

Source: ars-technica-ai · published 2026-04-28

TL;DR

• ~$162B in data center projects have been blocked or delayed since 2023, with $98B frozen in Q2 2025 alone.
• State audits show host states lose money: Virginia recoups ~48¢ per dollar of foregone sales tax.
• Opposition is cross-ideological — DSA + “Stop the Steal” in Michigan, Trump farmers + Food & Water Watch in Pennsylvania.
• Federal EO 14318 fast-tracks big sites but stops short of preempting local zoning.

The backlash has a dollar figure now

Ars frames rural opposition to AI data centers as a rising mood. The numbers are sharper than that. Data Center Watch tallies roughly $162 billion in project value blocked or delayed since 2023, with $98 billion frozen in Q2 2025 alone ¹³. That’s no longer a permitting nuisance — Equinix and Digital Realty have both cited deal-closure delays in guidance cuts, which means siting fights in Loudoun County and Newton County are now showing up in hyperscaler earnings models.

Litigation is producing wins, too. In August 2025, a Virginia circuit judge voided the 23-million-square-foot Prince William Digital Gateway because supervisors botched public-hearing notice requirements ¹⁴. The procedural template — challenge the meeting, not the merits — is being copied in other counties.

The fiscal pitch doesn’t survive an audit

The standard sell to a county board is jobs and tax base. Independent auditors keep finding the opposite.

Good Jobs First argues at least 14 states fail to disclose these losses in violation of GAAP ¹⁵. Meanwhile in Newton County, Georgia, homeowners ~1,000 feet from Meta’s facility describe wells producing “gritty sludge” and have spent $5,000+ on appliance repairs and filtration ¹⁶. Meta’s hydrology study denies causation — but no baseline well testing was done pre-construction, which is exactly the kind of detail that radicalizes a county commission.

A coalition the industry can’t lobby around

The political shape is what makes this durable. The Guardian documents Michigan “Stop the Steal” activists organizing alongside Democratic Socialists of America chapters against data center tax credits; in Pennsylvania, Trump-voting farmers partnered with Food & Water Watch to block an Amazon complex ¹⁷. Sanders and DeSantis have converged on moratorium-style proposals from opposite directions.

flowchart LR
    A[MAGA / property-rights right] --> C{Anti-data-center coalition}
    B[DSA / environmental left] --> C
    D[Farmers, well-water households] --> C
    C --> E[State moratoria<br/>ME 20MW cap, GA freeze]
    C --> F[Voided approvals<br/>Prince William]
    C --> G[Hyperscaler guidance cuts]

This is why the state-level bans passing — Maine’s 20 MW cap through 2027, Georgia’s pending freeze — are clearing legislatures of both colors. It’s not a partisan vulnerability that a single round of campaign donations can paper over.

Washington is hedging

The Trump administration’s response is split-track. Executive Order 14318 created “Qualifying Projects” status for any data center over 100 MW or $500M, granting FAST-41 federal permitting treatment to route around slow states ¹⁸. But the March 2026 “Ratepayer Protection Pledge” — signed by Meta, Microsoft, and OpenAI — commits developers to fund the full cost of grid upgrades rather than socializing them onto household bills ¹⁸. Notably, the framework backed off broad preemption of local zoning. Even the federal push respects how toxic siting has become.

The takeaway: the era when a county board would rubber-stamp a hyperscaler in exchange for a vague tax-base promise is closing. The audits are in, the wells are gritty, and the coalition fighting the next site doesn’t care who’s in the White House.

Round-ups

The missing step between hype and profit

Source: mit-tech-review-ai

MIT Technology Review’s Algorithm newsletter argues the AI industry is skipping the middle step between hype and profit, opening with a flyer picked up at an anti-AI march in London that echoed South Park’s underpants gnomes business plan.

Our commitment to community safety

Source: openai-blog

OpenAI publishes an overview of how it protects ChatGPT users, detailing model-level safeguards, automated misuse detection, policy enforcement workflows, and ongoing collaboration with outside safety experts and researchers.

Humanoid robots start sorting luggage in Tokyo airport test amid labor shortage

Source: ars-technica-ai

Japan Airlines is piloting humanoid robots at Haneda Airport to load cargo and clean aircraft cabins, part of a response to Japan’s chronic labor shortage in ground handling roles.

DeepInfra on Hugging Face Inference Providers 🔥

Source: huggingface-blog

DeepInfra joins Hugging Face’s Inference Providers lineup, letting developers route model calls through DeepInfra’s serverless GPU backend directly from the Hub or via the unified Inference Providers SDK alongside existing partners like Together, Replicate, and Fal.

Celebrating 20 years of Google Translate: Fun facts, tips and new features to try

Source: google-ai-blog

Google Translate marks its 20th anniversary with a retrospective post highlighting usage trivia and teasing new features for the service, which launched in 2006 and now spans hundreds of languages across text, voice, and live conversation modes.

Join the new AI Agents Vibe Coding Course from Google and Kaggle

Source: google-ai-blog

Google and Kaggle open registration for a June 2026 GenAI Intensive focused on vibe coding AI agents, extending the duo’s prior five-day generative AI bootcamps with hands-on agent-building exercises using Gemini models.

Quoting Matthew Yglesias

Source: simon-willison

Simon Willison quotes Matthew Yglesias saying that five months in, he’s decided he doesn’t want to vibecode himself — he wants professionally managed software companies to use AI assistance to ship better, cheaper products for him to buy.

Footnotes

1. Complete AI Training — GPT-5.5 release analysis — https://completeaitraining.com/news/gpt-55-releases-with-unverified-safety-stack-as-government/

   AISI red-teamers discovered a universal jailbreak within six hours… a configuration error in the provided testing environment meant AISI was unable to verify the effectiveness of the final fixes

   ↩ ↩² ↩³

2. ProPublica investigation into FedRAMP — https://www.propublica.org/article/federal-government-ai-cautionary-tales

   FedRAMP is a ‘rubber stamp,’ alleging that the program is severely understaffed and relies too heavily on assessments funded by the tech companies themselves

   ↩ ↩²

3. Rock Cyber Musings — ‘AI Attacker Advantage Is a Myth’ — https://www.rockcybermusings.com/p/ai-attacker-advantage-is-a-myth

   while AI can find bugs in seconds, the human-led process of patching, testing, and deploying fixes in complex enterprise environments still takes months… a ‘zero-day apocalypse’ risk where the lag between automated discovery and manual remediation becomes a permanent window of exploitation

   ↩ ↩²

4. Incrypted — investigation into OpenAI KYC vendor Persona — https://incrypted.com/en/investigation-on-openai-kyc-sparked-fears-of-mass-surveillance-of-crypto-users/

   Persona’s code contained mechanisms to transmit user data, including cryptocurrency addresses and biometric selfies, directly to federal agencies like FinCEN

   ↩ ↩²

5. Transformer News — https://www.transformernews.ai/p/openai-shouldnt-be-deciding-if-its-gpt-55

   OpenAI shouldn’t be deciding if its GPT-5.5 is safe enough to release

   ↩

6. Open Magazine — ‘#QuitGPT’ backlash coverage — https://openthemagazine.com/world/quitgpt-public-trust-fractures-as-openai-faces-backlash-over-pentagon-contract

   Caitlin Kalinowski, OpenAI’s former head of robotics, stated that ‘surveillance of Americans without judicial oversight and lethal autonomy without human authorization are lines that deserved more deliberation than they got’

   ↩

7. Korea Herald (Sovereign AI finalist selection) — https://www.koreaherald.com/article/10725712

   Naver Cloud and NC AI were eliminated after the first round, with expert panels citing a lack of technical ‘originality’ in Naver’s foundation model

   ↩

8. Seoul Economic Daily — K-Moonshot missions — https://en.sedaily.com/technology/2026/03/11/korea-unveils-12-national-missions-for-ai-driven-k-moonshot

   ₩10.1 trillion (approx. $7.27 billion)… a 206% increase over the previous year, with ₩464 billion committed between 2027 and 2031 to build independent foundation models in core sectors like biotechnology and materials

   ↩

9. Korea JoongAng Daily — Samsung/SK hynix LTAs — https://koreajoongangdaily.joins.com/news/2026-04-14/business/industry/For-Samsung-and-SK-hynix-longterm-deals-with-Big-Tech-offer-stability-in-churning-chip-cycles/2565305

   Samsung has reportedly secured a contract to supply more than 60% of Google’s HBM requirements for 2026… agreements include ‘price floor’ protections and a ‘prepayment deposit’ mechanism where Google pays between 10% and 30% of the total contract value upfront

   ↩

10. ResultSense — OpenAI for Countries comparison — https://www.resultsense.com/news/2026-01-21-openai-countries-global-ai-initiative/

    While Google focuses on research hubs, OpenAI has secured deals in the UAE and Norway to build massive data centers, effectively becoming the first customer for national ‘sovereign compute’ projects

    ↩

11. Korea JoongAng Daily — Hassabis/Lee Sedol reunion — https://koreajoongangdaily.joins.com/news/2026-04-29/business/tech/DeepMind-chief-reunites-with-baduk-champion-10-years-after-historic-match/2581021

    Hassabis reportedly estimated that Artificial General Intelligence (AGI) could be realized within the next five years… Lee Sedol warned that humans must not cede ‘sovereignty over thought’ or control of creative processes to machines

    ↩

12. Inkl — DeepMind staff revolt over Pentagon contract — https://www.inkl.com/news/incredibly-ashamed-google-deepmind-scientists-revolt-over-secret-pentagon-deal-to-use-ai-in-warfare

    over 600 employees protested a classified contract with the U.S. Department of Defense… researchers expressed ‘shame’ that the company removed previous pledges against military surveillance

    ↩

13. Data Center Watch (Q3-Q4 2025 report) — https://www.datacenterwatch.org/q3-q4-2025

    Approximately $162 billion in project value has been blocked or delayed since 2023, including $98 billion in the second quarter of 2025 alone.

    ↩

14. Virginia Business — Prince William Digital Gateway ruling — https://virginiabusiness.com/judge-voids-massive-prince-william-digital-gateway-project/

    A circuit court judge voided the 23-million-square-foot project, ruling that local supervisors failed to follow state advertising policies for public hearings.

    ↩

15. Good Jobs First — data center tax abatement disclosure report — https://goodjobsfirst.org/data-center-tax-abatements-why-states-and-localities-must-disclose-these-soaring-revenue-losses/

    Virginia’s JLARC found the state recoups only roughly 48 cents for every dollar of sales tax revenue foregone; Georgia auditors projected net state revenue losses up to $780 million by 2030.

    ↩ ↩² ↩³

16. Futurism — Meta Newton County water sediment — https://futurism.com/the-byte/woman-meta-ai-data-center-tap-water-sediment

    Homeowners roughly 1,000 feet from the Meta data center reported wells producing ‘gritty sludge’ and brown, sediment-heavy water; some families spent upwards of $5,000 on appliance repairs and filtration.

    ↩

17. The Guardian — ‘Datacenters: US political opposition’ (Jan 2026) — https://www.theguardian.com/us-news/2026/jan/13/datacenters-us-political-opposition

    In Michigan, a coalition formed between ‘Stop the Steal’ activists and the Democratic Socialists of America to protest tax credits for new data centers; in Pennsylvania, pro-Trump farmers worked with Food & Water Watch to block an Amazon complex.

    ↩

18. ACHR News — White House AI framework coverage — https://www.achrnews.com/articles/165992-white-house-ai-framework-could-boost-hvac-demand-as-data-center-growth-accelerates

    Executive Order 14318 created ‘Qualifying Projects’ status for data centers >100 MW or >$500M, granting FAST-41 treatment; the March 2026 ‘Ratepayer Protection Pledge’ commits Meta, Microsoft, and OpenAI to fund the full cost of grid upgrades.

    ↩ ↩²