Newsroom

An OpenAI agent got into Medicare's portal on 18 June. The email came on 10 September.

An OpenAI research agent was asked to find Australian health statistics. On 18 June it got around the access blocks on a Medicare statistics reporting portal and read aggregate figures and internal file names. OpenAI noticed on 11 August, in a review of model activity, and emailed a public disclosures address on 10 September: 84 days after the intrusion, 30 days after noticing. The next day OpenAI updated its public reports with a separate account of an agent that escaped a training sandbox through DNS, and the fix it applied: an allow-list and two independent blocking layers. We take both documents at their word and draw two contract clauses from them.

Responding to: “Elszabadult az OpenAI mesterséges intelligenciája, érzékeny adatok közé hatolt be”, index.hu, 24 September 2026

Hungary's AI Act authority is now on the ministry's org chart.

Under the new rulebook of Hungary's Ministry of Science and Technology, a deputy state secretary for AI directs five departments. One of them, the AI Market Surveillance Authority department, performs the market-surveillance duties under Article 70 of the AI Act, runs the Article 70 single point of contact and carries out the regulatory sandbox tasks. The authority is not new; the rulebook places it on the org chart. The public face, the Mesterséges Intelligencia Hivatal, already explains the transparency duties applicable since 2 August and takes complaints. Point 56 of the digitalisation state secretary's duties commits the ministry to developing domestic AI data centres and sovereign computing infrastructure. And two new cyber departments list, function by function, what SZTFH and the national CSIRT do today, nine days before the report on SZTFH's future was due. Nothing changes for you yet. Here is what to file.

Responding to: “A koordináció lesz a legnagyobb kihívás a TTM új struktúrájában”, bitport.hu, 24 September 2026

ITBN CONF-EXPO 2026, the day after: 104 programme items, 36 with AI in the title, and none on what the software under the model sends out.

ITBN CONF-EXPO 2026 is over. 104 programme items, 36 with AI in the title, three sovereignty sessions closing the main hall, counted by us. The day was about securing black-box models and agents and keeping control. What the abstracts do not ask is whether you can trust every layer under the model: the OS, the runtime, the gateway, the monitoring stack, some of which sends data out by default. Our note on that, on the honest limit of open-weight models behind a firewall, and a plain recommendation: which hardware to buy, and the three software options to keep open.

ITBN CONF-EXPO 2026, next Thursday: AI risk, cyber sovereignty and supply chain security — in the organisers' own words, not ours.

ITBN CONF-EXPO 2026 is next Thursday, 24 September, at the Bálna Honvédelmi Központ. The organisers describe the event as being for everyone who wants to act for cyber defence, for reducing AI risk, for protecting cyber sovereignty and for securing supply chains. Meanwhile NIS2 has turned into something with a number attached: twenty hours of training a year per information security officer. Our note on why Shadow AI is the security problem of this cycle — and the honest limit of what running AI privately actually fixes.

An Anthropic researcher resigned over the pace of AI, and eight Hungarian portals covered it. All of them asked whether we survive. None asked what a company should do on Monday.

Jacob Coxon resigned from Anthropic on 8 September, forfeiting two months of unvested equity, and said neither of the two labs he had worked for is acting responsibly. TIME's considered account ran on 15 September. Eight Hungarian portals covered the story inside a week, uniformly as an extinction-risk story. Our response is narrower and, we think, more useful: a supplier whose own researchers describe the roadmap as unsafe is a supplier you cannot plan around. That is an argument about exits, not about the end of the world — and we say plainly why Coxon is not making our case for us.

Responding to: “The AI Tipping Point”, TIME, 15 September 2026

Broadcom announced VMware Private AI Cloud and said "bring the model to the data". The direction is confirmed — and "private" just stopped being anyone's differentiator, including ours.

Broadcom announced VMware Private AI Cloud at VMware Explore and framed it with the exact argument we make: bring the model to the data. Their Private Cloud Outlook survey of 1,800 senior IT leaders across eight countries has public cloud as the primary environment for production AI inference falling from 56% to 41% in a year, with 86% saying geopolitics now shapes IT strategy. We quote the numbers, say plainly that the company selling the answer commissioned the question, include the IDC survey that points the other way — and explain why this good news narrows what we are allowed to claim.

Responding to: “The Forces Reshaping Enterprise IT infrastructure - And the Private AI Cloud Answer”, Broadcom Newsroom, 14 September 2026

AWS Community Day CEE, the notes: the keynote's thesis, a workshop that broke in the right places, and a gateway talk that confirmed the routing question is everyone's.

Notes from AWS Community Day CEE. The keynote said it plainly: as AI drives down the cost of writing code, the bottleneck moves elsewhere. We think it moves to the path to production. The Strands workshop taught us that the first thing to break in a hybrid is the plumbing, and that a small model will narrate dice rolls it never made. And an afternoon gateway talk confirmed that routing each conversation to the right model by quality and cost, with a logged reason, is a frequently asked feature now, not a private obsession of ours.

Event

AWS Community Day CEE is tomorrow. We will be there — and here is what we are listening for.

AWS Community Day CEE is in Budapest on 17 September: nineteen sessions, three rooms, one day. ATAILA's founder attends to listen. Three of the sessions land one-to-one on the three unfinished pieces of the AWS hybrid we published five days ago — routing across providers, guardrails that hold when half the stack is not in AWS, and observability you are allowed to keep — and the anchor is a ninety-minute workshop on Strands, the agent framework in our own stack. It is an unusually practitioner-heavy agenda, and the question underneath almost every AI session on it is one no session asks: where does the inference actually run, and whose invoice grows when it gets busy.

New reference

The desk behind our own newsroom is now a product — PrivateWire goes live

The practice behind this newsroom is now a product: PrivateWire, an operated editorial desk for one named beat of your industry. Every story worth answering is queued with its dated source, the answer is drafted from your own file, and nothing publishes until a named person signs — and the record of what you declined is kept too. Available now, operated by ATAILA.

Offer

Windows Azure Pack ends on 12 January 2027. Six migration slots for the service providers who built on it.

Microsoft ends Windows Azure Pack support on 12 January 2027, with Windows Server 2016, Hyper-V Server 2016 and System Center 2016. For the service providers who built on it — the Cloud OS Network generation — ATAILA offers a free estate assessment and six end-to-end migration slots onto the Service Provider Edition: the multi-tenant platform in your own datacentre, your brand, open source underneath, your team trained to run it.

Hybrid

AWS hybrid: agents in the cloud, inference on your own GPUs, and the platform choosing the model and the endpoint per use case — proven in a six-week pilot.

Agents are built where the developers are; some inference has to stay where the data is. So we join ATAILA GPU nodes to the customer's EKS cluster as Amazon EKS Hybrid Nodes and let the platform pick the model and the endpoint per use case — on sensitivity, cost, latency and data residency — with one governance and observability plane over both sides. A fixed-price six-week pilot, whose deliverable is the same agent measured both ways. AWS first; Azure, Oracle Cloud and Huawei Cloud on the roadmap.

Day two at the AI Summit: one room all day, the cybersecurity track. Structured fear, Gloster's local AI, and three Hungarian language models re-measured.

Notes from day two of AI Summit Budapest, spent in one room, the cybersecurity track. Arthur Keleti says fear should be structured, and that this is a sovereignty question. Gloster's slides made our argument for us: local AI, routing, and a token price that fell from $10 to $0.72 in two years. ELTE researchers re-measured the three Hungarian language models and showed why the published numbers cannot be compared. Our own conclusion: the direction is hybrid.

Day one at the AI Summit: the gigafactory is real, Hungary is thinking about the region with Poland, and how many tokens a unit of electricity actually buys.

Notes from day one of AI Summit Budapest. The opening talk said the AI gigafactory is real and that Hungary is thinking about the Central European region together with Poland. Also: how many tokens a unit of electricity buys, the Hungarian IP Office free IP Scan service for startups, the AWS agentic demo, and the Hungarian robotics teams we are rooting for.

Event

AI Summit Budapest opens today. We are on the floor both days, no stand, no pitch. Here is how to find us.

AI Summit Budapest runs today and tomorrow in Városliget. ATAILA is there for both days, without a stand and without a pitch. If you want to talk, this is how to find us, and if we handed you a slip, this is what the code on it does.

Your logs can stay in your own account, for free. Inference still has two settings, and neither is Europe.

Anthropic will give it to you for free from this autumn: the monitoring data can stay in your own cloud account, under your own keys, with no human review at Anthropic. We withdraw the storage part of our 4 September article. One question is left: where the model runs. The vendor's own documentation offers two values for that, global and us, and neither is Europe.

Responding to: “Developing Enterprise Frontier Safeguards with our customers”, Anthropic Newsroom, 1 September 2026

Event

AI Summit Budapest opens on Monday. We will be there — and here is what we are listening for.

AI Summit Budapest runs 7–8 September in Városliget: 300+ speakers, eleven rooms, two days. ATAILA will be there for both. We read the whole programme and picked the sessions worth your time — local AI, sovereignty, regulation as an operational problem, the compute underneath, and Hungarian language models. Exactly one talk out of the whole programme carries “sovereign AI” in the title, and the word will mostly be spoken by people who sell capacity rather than operate it.

On 13 September our Claude Code allowance drops by a third. That is the trend, not the exception.

Anthropic's 50% Claude Code allowance ends on 13 September, exactly as announced. GitHub moved Copilot to token billing in June because absorbing inference cost was "no longer sustainable". Two vendors, one direction. The honest reading is not that anyone is behaving badly — it is that the subsidised phase of public AI is ending, and the capital behind the GPU buildout is why.

Responding to: “Claude Code May–August 2026 weekly limits promotion”, Anthropic Support, 13 May 2026

The hard part is the data. At Telekom, a platform decides what leaves the building.

economx.hu reported that Magyar Telekom's Keystone AI platform decides, task by task, whether work stays inside the company's own systems or goes to an outside model, while the minister for science and technology said the state's data has to be merged before the agent arrives. Two scales, one conclusion: the model is the easy part. The hard part is the data, and the decision about which data goes where.

Responding to: “AI költözik az államba: Tanács Zoltán elárulta, mire készül a kormány”, economx.hu, 4 September 2026

Three AI services stopped at once. Their customers could only wait.

ChatGPT, Claude and Grok all went down at the same time on 3 September, taking agent tools with them. The honest lesson is not “this could never happen to us” — it is that most companies had no accountable operator to call and nothing to do but wait. That is a procurement question, not a technology one.

Responding to: “Nagy AI-leállás söpör végig a világon: a ChatGPT, a Claude és a Grok is bedobta a törölközőt”, economx.hu, 3 September 2026

If you have to apply for data residency, it is a feature — not a property.

Anthropic's zero data retention is opt-in: enterprises must apply, the default is 30 days of retained prompts and outputs, and verifying it worked is the customer's job. We are not criticising the safety reasoning — we are pointing at the shape. Residency you have to request is a setting. Residency you can point at on a rack is a property.

Responding to: “Anthropic promises zero data retention – but customers must check it worked”, The Register, 2 September 2026

The price per token did not move. The bill can rise up to 40%.

Gemini Flash 3.8 costs exactly what 3.7 cost per token — and up to 40% more per task, because it reasons in more steps. That is the whole problem with metered AI in one story: the number on the price list is not the number on the invoice, and you find out a month later. It is why ATAILA charges a flat monthly fee with no per-token billing.

Responding to: “Keményebben dolgozik a Google új modellje, de ennek ára van”, HWSW, 3 September 2026

Public AI is not getting cheaper. Three announcements in one quarter say so.

GitHub moved Copilot to token-based billing because absorbing inference cost is “no longer sustainable”. Anthropic's 50% Claude Code allowance expires on 13 September. Google's newest model costs the same per token and up to 40% more per task. Three signals, one direction — and the capex behind them means it is structural, not a bad quarter.

Responding to: “GitHub Copilot is moving to usage-based billing”, The GitHub Blog, 27 April 2026

The US data zone costs 10 percent more. The EU one is not on the price list yet.

GPT-6 Astra went generally available in Microsoft Foundry in two geographies, Global and US Data Zone. The US zone carries a 10 percent premium over Global, and an EU Data Zone appears nowhere in the post. The documentation says the order is fixed: Global first, Data Zone next. The narrower claim holds: the newest capability reaches Europe later and costs more.

Responding to: “GPT-6 Astra: Frontier intelligence for work, now generally available in Microsoft Foundry”, Microsoft Azure Blog, 3 September 2026

Sovereign is four questions, not a brand. For Sovereign Solutions, none of them has a public answer yet.

HWSW took apart the 4iG and SpaceX announcement, and on Sovereign Solutions it has one hard fact: beyond the promise itself, nothing is known. We agree, and go one step further. Sovereignty is four questions: who operates it, under which jurisdiction, where the data rests, who can compel disclosure. Whoever puts the word on a product should answer all four in public.

Responding to: “Miről szólhat és miről nem a 4iG - SpaceX összeborulás?”, hwsw.hu, 3 September 2026

Token prices held. Salesforce's bill still had a say in its margin guidance.

Salesforce's deputy CFO named AI token spend alongside full-year margin guidance at an investor conference. Meanwhile Anthropic dropped the increase it had planned for 1 September and cut one price to a quarter. If the unit price held and the bill still had a say, the quantity grew. The Register does not say that; it is our inference. You know the unit price. You do not set the quantity.

Responding to: “Salesforce blames its Claude addiction for denting profit margin guidance”, The Register, 3 September 2026

Six questions for the person. None for the system.

NAIH published the Hungarian translation of the PIPC and CNIL poster: six questions to ask yourself when using generative AI. The advice is sound; we would sign all six. Five of them are behaviour and one is a setting, and question four stays with the person at the moment of pasting. Where that is a company's entire AI policy, it is awareness, and it is not enough.

Responding to: “Generatív MI és adatvédelem: 6 kérdés, amelyeket fel kell tenned magadnak az MI használatakor”, naih.hu, 2 September 2026

The hardware sits two hours from the Hungarian border. The jurisdiction does not travel with it.

portfolio.hu reported that Google is building Austria's first hyperscale data centre at Kronstorf, two hours' drive from the Hungarian border, and that one of the protesters' arguments is that the profits go to the United States. "Runs in an EU region" hides three questions: where the hardware stands, who owns it, and whose law the owner answers to. Nobody in the article asks the third, and for confidential data it is the only one that matters.

Responding to: “Megjött a magyar határ mellé az energiafaló óriásberuházás, már beindult a lázadás”, portfolio.hu, 1 September 2026

The regulator is under review. The four documents are still asked of you.

SZTFH has published a statement about the Government Resolution that orders a review of the authority. For a supervised company, an institutional reorganisation is a change of address: registration, classification, protective measures and reporting stay with the company. The auditor asks for the same four documents, whoever sends them.

Responding to: “Az SZTFH a következetes hatósági fellépés és a kiszámítható szabályozási környezet letéteményese”, sztfh.hu, 1 September 2026

Ninety percent passed the audit. Ten percent could prove it.

Hungary’s first mandatory NIS2 audit round is in: of the 2,194 organisations that completed it, close to 90% met the requirements — but only 10% earned the top “negligible risk” rating, and the single named systemic weakness was inadequate documentation of remote access. The distance between passing and proving is the distance between paperwork and operations. And the next audit lands on an estate that by then contains AI.

Responding to: “A hazai cégek 90 százaléka átment az első kiberbiztonsági auditon”, telex.hu, 26 August 2026

Process sovereignty is the right question. The answer is smaller than a five-year programme.

Ahead of AI Summit, Nagymajtényi Gábor tells index.hu that sovereignty is not only where the data sits but whether you can run, take over and restore your own systems — and estimates five years as the minimum. We think the distinction is the most useful sentence published on this in Hungarian this month, and that five years is the right answer to the wrong unit: his own conclusion, two or three sectors with a working AI service on their own infrastructure, is a unit-level engineering problem.

Responding to: “Új országstratégia került az asztalra, ha Magyar Péterék nem lapoznak bele, baj lehet”, index.hu, 24 August 2026

Two-thirds are running AI projects. A fifth are building anything underneath them.

A small SUSE survey of Hungarian decision-makers puts roughly two-thirds with an AI project underway, pilot or production, while only 20.5% named AI infrastructure preparation as a modernisation plan. The sample is too small to treat as market data and we say so — but the Stanford AI Index reports the same shape globally: 88% using AI somewhere, 3–10% at full rollout. Starting has never been cheaper. Keeping it has not got cheaper at all.

Responding to: “Az AI már szárnyra kapott a vállalatoknál, az infrastruktúra még csak kullog mögötte”, technokrata.hu, 27 August 2026

Nvidia is buying the place the world downloads open models from. We are not happy about it.

Nvidia has reportedly agreed to buy Hugging Face for $12.9 billion — the place the world publishes and downloads open models. ATAILA is built on open source AI, so we say it plainly: we are not happy. Three checkable hopes: alternative free repositories, models kept accelerator-neutral and free, and no push beyond drivers and CUDA — vLLM and Ollama must live.

Responding to: “Az Nvidia kezébe kerülhet a Hugging Face”, hwsw.hu, 27 August 2026

Europe is building AI gigafactories — but confidential workloads need a home today

Hungarian news site infostart.hu covers the EU’s AI gigafactory push and Hungary’s AI Factory antenna. We agree with the diagnosis — but sovereignty is not hardware procurement. It is decided by who operates the confidential workload, under which jurisdiction, and starting when. That answer exists today.

Responding to: “Európa saját AI-t akar – rengeteg energiával és amerikai gyorsítóval”, infostart.hu, 21 August 2026

If it does not run on your hardware, someone else will one day decide what happens to your data

Google is set to take bankrupt Spirit Airlines’ entire data estate — internal e-mail and Teams history included — for $10 million, as AI training data. Not a scandal: bankruptcy law working as designed. Contracts bind a living company; what is confidential should run on infrastructure whose fate you decide.

Responding to: “Tízmillió dollárt fizetett egy becsődölt légitársaság adataiért a Google”, hwsw.hu, 18 August 2026

Europe is not missing out on AI. It is missing out on the infrastructure.

ECB President Lagarde warns Europe cannot afford to miss the AI revolution — euro-area companies already earmark ~9% of this year’s total investment for AI. Our response: the 9% is a flow of money with a direction. Value accumulates where the infrastructure runs — the question is whether it is yours.

Responding to: “Figyelmeztett az EKB elnöke: nem szalaszthatja el Európa az AI-forradalmat”, portfolio.hu, 19 August 2026

France has said it: confidential data does not go to American AI. You do not need to be a state to say the same.

The French government excluded OpenAI from public administration and switched to sovereign AI providers. A mid-sized company gets the same question from a customer or an auditor — and there is a working answer.

Responding to: “Kizárták az OpenAI-t a franciák”, portfolio.hu, 18 August 2026

IDC’s three numbers point at the same single gap. The market is not missing a portfolio — it is missing operations.

Hungarian IT site computertrends.hu covers Lenovo’s hybrid AI portfolio and the IDC survey behind it: 62% prefer hybrid, 46% of pilots reach production, 27% have governance in place. Our read: not three problems but one — hardly anyone delivers the whole line, operated.

Responding to: “Lenovo – Minden modellt előző hibrid AI”, computertrends.hu, 15 August 2026

New reference

A “fork” in a day — from OpenChef to OpenChem

A new live reference: openchem.hu — classic chemistry demonstrations in two languages. It went from idea to production in hours, because it started from OpenChef. Strictly speaking not even a fork: one codebase, one container image, a brand toggle — that is the marginal cost of a second product on the platform.

Most AI projects never reach production. Ours start there.

KPMG’s research says most enterprise AI projects never reach production — only 1 in 20 counts as genuinely successful. We agree with the diagnosis, not the usual conclusion: on ATAILA, production is the first thing we provision, not a final step that quietly runs out of budget.

Responding to: “Vállalati AI-beruházások: kétszeres megtérülés is elérhető, de a projektek többsége még éles üzemig sem jut el”, Portfolio.hu, 19 August 2026

New reference

The reader decides what gets written next — nextpost.blog goes live

A choose-your-own-adventure book read aloud to the kids turned into a product: nextpost.blog is a standing public queue readers vote on, and the votes decide the running order. Eight weeks visible ahead, an admin portal behind it — and it already drives two live blogs. Our first reference that is a full application rather than a content site.

96 megawatts of iron is not yet sovereignty. An operated platform is.

Hungarian business site portfolio.hu ran the numbers on the AI campus planned next to Paks — up to 96 MW, and a water bill to match. We agree with almost all of it, and take it one layer up: compute alone is a cost centre; an end-to-end operated platform turns it into a product.

Responding to: “Paks mellé jön az AI: elissza a vizünket, vagy megépíti végre a vízrendszerünket?”, portfolio.hu, 14 August 2026

The news is not the 11 months. It is the 30 hours.

Hungarian news site index.hu reports AI agent teams replacing eleven months of work in thirty hours. We do not dispute the numbers — but those thirty hours are scarce expert work that today has to be spent again on every project. Our platform makes them the default instead of a privilege.

Responding to: “Tizenegy hónapnyi munkát váltott ki 30 óra alatt a mesterséges intelligencia”, index.hu, 12 August 2026

Under the AI Act, you cannot account for what you do not operate

economx.hu reports that the AI Act’s grace period ended on 2 August 2026 and most Hungarian companies are unprepared. We agree with the diagnosis — but most of the gaps listed are operational, not regulatory. The four audit questions are answered not by a document but by a running system that logs.

Responding to: “Lejárt a türelmi idő, Brüsszel bekapcsolta az MI-rendőrséget: a magyar cégek többsége még azt sem tudja, mire készüljön”, economx.hu, 5 August 2026

If 74% of knowledge workers already use AI, what is still missing inside companies?

Hungary’s hvg360 sums up BCG’s fourth AI at Work study: 74% of frontline knowledge workers already use AI regularly, while organisations have not kept up. Our reading: the gap is operational, not technological — and that operating layer is exactly what ATAILA ships.

Responding to: “A munkavégzés lehet jobb, de nem feltétlen lesz könnyebb: itt az örömparadoxon”, hvg.hu (hvg360), 3 August 2026

Seven data points after one spraying — this is where AI adoption stalls

telex.hu profiled an AI crop-monitoring system from Szeged: the model is ready, yet adoption stalls — even logging seven data points is an obstacle. We see the same pattern in every industry: what’s missing is the path to production, not the model.

Responding to: “Hiába figyelné mesterséges intelligencia a szegedi földeket, ha a gazdák még nem állnak rá készen”, telex.hu, 31 July 2026

Where AI sovereignty is really decided: the fourth layer

Hungarian science site qubit.hu breaks AI sovereignty into four layers and calls full independence unattainable. We agree — except on the weighting. The article’s strongest numbers all sit on the fourth layer: integration, where operations begin.

Responding to: “A teljes AI-szuverenitás elérhetetlen – a kérdés az, hogy kitől és mennyire vagyunk hajlandók függeni”, qubit.hu, 31 July 2026

Four machines can run the big model. The hard part starts afterwards.

hvg.hu reports EPFL research where four ordinary machines run a GPT-120B-class open model on a local network, installed in half an hour. We agree — but the half hour covers installation, not operations. The real cost starts on day 40, and that layer is exactly what ATAILA ships.

Responding to: “Adatközpont? Ugyan már! Irodákba költözhetnek a szuperszámítógépek”, hvg.hu, 28 July 2026