🤖 EU AI & Sovereignty Top 10
8 November 2026
This weekend’s edition has two tracks running in parallel. On the financial side, the ECB, the ESM and the IMF all warn that a correction in US AI stocks would hit Europe harder than Brussels admits — in the same week that the EU’s first ever AI summit in Dublin focuses only on uptake and human control. On security, the British NCSC calls the recent fridge attacks the “foothills” of a new reality, and a benchmark shows open Chinese models nearly matching closed frontier models at offensive cyber. Sovereignty, meanwhile, remains a paper promise: the UK is putting a roughly £1 billion AWS contract on the market and the NHS has to ask Palantir to prove its own success claims.
financiën
🔥 HOOG
‘A tale of two bubbles’: Brussels inflates the AI bubble while Dublin looks away
Rapporteur sets next week’s leaders’ summit in Dublin — the first EU summit ever devoted entirely to AI — against the financial risks the EU’s own institutions are describing. The two priorities of Council President Costa, more AI uptake and keeping AI “under human control”, do not touch the actual problem: Europe’s exposure to a correction in US AI shares. US market capitalisation is now about 230% of GDP (against roughly 150% before the dotcom crash), the US share of European foreign equity holdings rose from 20% in June 2022 to over 30%, and the ECB estimates European households hold some €440 billion in US tech stocks without being aware of that concentration risk. Why it matters: ESM, ECB economists and the IMF all warned this year about “abrupt repricing” — and this week’s market slide is being read as the first sign.
cybersecurity
🔥 HOOG
The Guardian: the fridge attacks are the NCSC’s “foothills” of a new cyber reality
The Guardian ties ten days of incidents into one picture: the fridge and freezer attacks at supermarkets in Belgium, Estonia, Ireland, Finland, Luxembourg, the Netherlands and Sweden (temperatures up to 50 °C), the attack on French blood banks, the $6.3 billion Binance heist of 30 October and the false nuclear alert in Ukraine’s Air Alert app on 1 November. The cooling equipment was reached through the vendors of its control systems, some end-of-life with factory default passwords from the manual. NCSC head Richard Horne calls this the “foothills”: the biggest short-term risk is not one large attack but a huge increase in the number of incidents. Since Anthropic released its Mythos model in April, vulnerabilities found by major tech firms have risen eightfold. Why it matters: Finland, with 1,340 km of border with Russia, logged almost 21,000 cyber incidents last year against about 13,000 in 2022 — and the attacks carry Russian fingerprints without being firmly attributable.
soevereiniteit
📌 MEDIUM
UK puts a roughly £1 billion Home Office cloud deal to AWS on the market
A procurement notice from 30 October values the renewal of the Home Office’s public cloud services at £881.6 million excluding VAT — well over £1 billion including it, against roughly £450 million for the contract AWS won in 2023. The deal runs from December 2027 to December 2029, with an option to 2031. BidSkim reckons AWS already takes 75% of all UK government cloud spend (about £1.22 billion) against £355 million for Azure. Set that against a survey in which about three quarters of companies in the UK, France and Germany fear Washington could suddenly cut off access to US technology platforms. Why it matters: this is not an EU decision, but it is the sharpest illustration of the pattern the EU’s Cloud and AI Development Act is meant to break — lock-in that weakens the buyer’s negotiating position, as an internal Cabinet Office note already admitted in 2024.
aanbesteding
📌 MEDIUM
NHS forced to ask Palantir to substantiate its own success claims
NHS England chief executive Jim Mackey told the Health and Social Care Committee this week that he will write to Palantir to have future claims about the benefits of the Federated Data Platform cleared with NHS England in advance — while conceding that this is hard to enforce contractually. The reason: Palantir executive Tom Watson keeps repeating that 137 of 139 trusts report benefits, but omits the caveat that there is no causal study; an FOI request earlier showed that nearly a third of trusts using the platform actually performed fewer procedures than before. The £330 million contract from November 2023 runs out in March 2027; the independent evaluation by Imperial College Projects, awarded in February 2026, is only expected after the renewal decision. Why it matters: same pattern as the cloud — a US vendor deep inside public infrastructure, with a weak contractual grip for the buyer.
ai-veiligheid
🔥 HOOG
WIRED: open Chinese models almost match closed frontier models at hacking
Vals AI let models break into simulated corporate networks on their own, without help and without prior knowledge. OpenAI’s Daybreak model, available only to trusted partners, reached 81% of the targets; China’s open-weight DeepSeek reached 73% — equal on some tasks. The price gap is larger than the quality gap: $17 to $46 in tokens per attack for GPT-6-Sol against $2 to $3 for DeepSeek. That undermines the frontier labs’ strategy of gated access to slow the spread of offensive capability: once weights are public, no provider can restrict use. Why it matters: researchers note that real criminals still prefer cheaper older models, but for defenders the argument for using open models is now stronger than the argument against.
onderzoek
📌 MEDIUM
AI auditors catch almost everything but drown reviewers in false alarms
Andy Hall of the Anthropic Institute replicated METR’s MALT dataset and had five models detect human-labelled AI agent misbehaviour. At a 5% false positive rate, GPT-5 caught almost 98% of reward hacking cases while Gemini 3.8 Flash barely reached half; on sandbagging all auditors caught 71 to 81%. Newer and more expensive models are not consistently better auditors — GPT-5 beat GPT-6 Astra on reward hacking at about an eighth of the cost, and the false alarm rate barely falls across generations. In his worked example of 100,000 agent sessions per month, a 3% false positive rate produces almost 3,000 false alerts against about 2,000 true hits. Why it matters: for AI Act enforcement, audits can estimate an incident frequency but are not yet a reliable gatekeeper per event — and reviewers who learn to ignore alerts are how real incidents slip through.
investeringen
📌 MEDIUM
Europe’s quantum startups get mostly private money — and that is the problem
According to McKinsey’s Quantum Technology Monitor 2026, investment in quantum startups reached €11.15 billion in 2025, 6.3 times more than a year earlier, but only 3% came from public sources — governments, sovereign wealth funds, universities — against about a third in 2024. That sits alongside big public quantum programmes, but that is different money: research, lab and test infrastructure versus capital on a company’s balance sheet. Europe’s base is strong: the EU produces 22% of the world’s quantum computing publications and 17% of the top 10% most cited papers, and Germany hosts 12 of the world’s 69 quantum master’s programmes. Why it matters: research funding alone does not carry the scale-up phase, and quantum is one of the few fields where Europe leads technically — exactly the position that can be lost through US capital or acquisitions.
defensie
📌 MEDIUM
TEKEVER raises €520 million and plants its defence-tech flag in the Nordics
The Portuguese defence-tech company closed a €520 million Series D at a €5.7 billion valuation; Denmark’s state investment fund EIFO put in €13 million and TEKEVER announced Nordic expansion at the same time. It is one of the largest European defence-tech rounds to date and fits the broader pattern of this week: European capital, public and private, is visibly shifting towards defence and dual-use technology. Why it matters: that is precisely the domain where sovereignty and AI procurement meet — drones and autonomous systems run on European money and, increasingly, European models.
begroting
📌 MEDIUM
Costa tries to unblock the EU budget with a narrower CORE levy
Council President António Costa proposes limiting the contested CORE own-resources levy to companies with annual turnover above €750 million, and turning it into a 0.05% levy instead of a fixed amount per company, as the Commission proposed last year with a €100 million threshold. The concession is meant to bring Germany and like-minded member states on board, but two diplomats doubt to POLITICO that it is enough. The total revenue forecast of €65 billion a year is unchanged — a concession to Paris. Costa tables a new negotiating document on 21 November ahead of a special summit. Why it matters: this is the same MFF knot that holds AI chip financing and the Horizon budget hostage — as long as own resources are unresolved, the digital and research budgets stay frozen.
open modellen
📌 MEDIUM
Chinese state propaganda in Qwen turns out to be a thin layer that fine-tuning strips away
Nous Research fine-tuned Alibaba’s open-weight Qwen3.8-27B with 53 million tokens of supervised fine-tuning and cut the share of answers containing Chinese state propaganda from 22% to under 1%, without losing general capability. Two implications for the European debate: alignment in Chinese open models is technically reversible, so “open Chinese model” is not the same as “Chinese government line in your product”; and once a model is downloaded there is no provider left who can withdraw or audit its behaviour. Why it matters: that is exactly the question the AI Act tries to answer through transparency and documentation duties — not only what a model does, but who is still responsible once the weights circulate freely.
📈 Key Trends
- • Financial exposure replaces capability as the main risk. ECB, ESM and IMF all point at the same thing: the danger is not that Europe cannot build frontier models, it is that European households and pension funds are long on them.
- • Sovereignty is negotiated in contracts, not in communiqués. A £1 billion AWS renewal and a £330 million Palantir deal show where the leverage actually sits — and it is not with the public buyer.
- • Open weights undercut gated access. 73% versus 81% at a tenth of the price means export-control-by-API no longer holds as a strategy, for offence or defence.
- • Incident volume, not incident size. Fridges, blood banks, a false nuclear alert: the NCSC “foothills” framing is about frequency, and AI tooling is accelerating vulnerability discovery eightfold.
- • Oversight tooling is not yet oversight. Auditors that catch 98% but generate 3,000 false alarms per 100,000 sessions cannot carry AI Act enforcement on their own.
Wat vond je van dit bericht?
