
🤖 EU AI & Sovereignty Top 10
20 November 2026
The through-line today is that the EU’s own executive is now a case study in its own policy. The Commission confirmed an AI-assisted cyber campaign against its systems that lasted ten days, days before digital ministers meet on AI safety. Around that: Costa floats a digital tax on Big Tech to unblock a €2 trillion budget — a levy that became thinkable only because Washington is distracted; Politico documents how AI attack tooling is outrunning the volunteers who maintain the open web; UK AI minister Narayan says no to a British sovereign frontier model; startups at Slush fight over EU Inc; Brussels moves from content moderation to ranking regulation against TikTok; the UK safety institute breaks GPT-6.2 Astra with half the queries of its predecessor; and Washington’s export regime turns out to have holes on both ends — licences for army-linked Chinese buyers and a cloud loophole closing only now. Anthropic, meanwhile, closes its first day near $1.7 trillion while the rest of the market sinks. Nine publications, no overlap with the editions of 17, 18 and 19 November.
cybersecurity
🔥 HOOG
The Commission confirms an AI-assisted cyberattack on its own systems
The European Commission confirmed on Thursday that a cyber campaign has hit several of its services since late October. Spokesman Thomas Regnier described it as “a series of incidents” and confirmed an AI tool was used; Politico reports it was ARTEX, an open-source AI that hunts vulnerabilities in IT systems. Detection was on 22 October and, according to Neowin (citing Reuters), the intrusion ran for ten days, taking out HR systems for payslips, time registration and medical reimbursements. The Commission says it has no evidence that staff data was stolen and will not estimate the scope while the investigation runs. Note the discrepancy: Neowin reports that large amounts of personal information were exposed, the Commission denies it. Why it matters: the EU’s own executive just demonstrated both halves of the argument at once — AI as attack tool and AI as defensive tool — weeks before digital ministers meet on AI safety on 1 December.
financiering
🔥 HOOG
Costa floats a digital tax on Big Tech to close the EU budget
European Council President António Costa is exploring whether he can get government leaders behind a tax on tech companies such as Google and Meta, to unblock the roughly €2 trillion EU budget and the closing deals around it. Four EU diplomats and officials tell Politico that the poor Republican showing in the midterms has reduced the threat of US retaliation — until recently the main reason to leave such a levy alone. Costa presents a revised 2028–2034 budget on Saturday and is aiming for agreement before elections in France, Italy and Poland. His team is considering replacing the contested CORE corporate levy with a digital one; Germany has warmed to it, Ireland — home to much of American tech — is against. Why it matters: a digital levy is the one instrument that would make European tech policy pay for itself, and it is now being discussed for budgetary rather than industrial reasons. That is both why it might pass and why it might be designed badly.
open-source
🔥 HOOG
AI is finding the open web’s holes faster than volunteers can patch them
Politico spoke to more than a dozen cybersecurity leaders and open-source maintainers: last month’s wave of outages and attacks comes from new AI tooling that locates vulnerabilities in existing code far faster than humans can close them. “We’re running on fumes,” says Robin Bender Ginn of the OpenJS Foundation; Helmut Reisinger, Europe head of Palo Alto Networks, sees “more or less” ten times as many incidents as four months ago and expects that to become “the new normal”. The carriers are small non-profits — PHP Foundation, OpenJS, Python Software Foundation — with a handful of paid staff funded by donations and government support. When the recent PHP holes were patched, the person who had to ship the updates was simply unavailable because of a scheduled medical procedure. Why it matters: this is the least glamorous dependency in the stack and the one an AI-driven attack wave exposes first. Every sovereignty plan about models and chips quietly assumes somebody is still maintaining the plumbing for free.
soevereiniteit
📌 MEDIUM
UK AI minister rejects calls for a British sovereign frontier model
UK AI minister Kanishka Narayan rejects calls to build a British sovereign frontier model, and according to Sifted that drew angry reactions from the sector (“I’m so angry”). The dispute goes to the core of the European sovereignty file: is owning compute and model development a strategic necessity, or is a country better off leaning on foreign labs and organising only oversight and adoption? Limitation, stated plainly: Sifted sits behind a cookiebot/DataDome wall — the body was not readable, so this item rests on the headline and subheading only. Why it matters anyway: London is the reference case Brussels keeps citing in both directions. If the UK concludes that a sovereign frontier model is not worth the money, that is a usable argument for the “applications, not labs” camp — and it will be quoted in the CAIDA and AI Act debates within days.
startups
📌 MEDIUM
At Slush, startups demand Brussels withdraw or rewrite EU Inc
Startup lobby groups asked Commissioners Henna Virkkunen (tech) and Ekaterina Zaharieva (startups) at Slush to withdraw the EU Inc proposal and refile it. In their reading, member states in the Council have hollowed out core parts such as the freedom to incorporate in any member state and simplified insolvency rules. “This is a chance for member states to reconsider their position,” says Simon Schaefer of Startups for Europe. General Catalyst’s Jeannette zu Fürstenberg attacks the “dry income” levy on illiquid shares: “you are bankrupting the people who build it.” A separate group of 38 tech executives (Daniel Ek, John Elkann, Niklas Zennström) takes a softer line, calling EU Inc “the most important first step” and asking Brussels to keep enforcing the DMA alongside it. Why it matters: the EU’s flagship attempt at a single company law is being fought over by the people it was written for, and the split is between those who want a better text and those who want the text to survive at all.
platform-regulering
🔥 HOOG
Brussels turns the screws on algorithmic rabbit holes — TikTok preliminarily in breach
The Commission is moving towards binding rules for recommender systems themselves. Earlier this month it preliminarily found that TikTok’s “For You” algorithm pushes users — minors in particular — towards harmful content including self-harm and suicide material, which in Brussels’ reading breaches the Digital Services Act and requires the app’s core function to change. TikTok rejects the findings and points to users’ “direct and intuitive control”. Researchers at CCDH created six accounts as 13-year-old boys this week, liked “looksmaxxing” content and were served self-harm content within five minutes of scrolling. Why it matters: this DSA case is explicitly the overture to the upcoming EU child-safety rules, where requirements on recommender systems are the centrepiece. Regulating the ranking rather than the content is the first genuinely structural obligation in the DSA, and it is the one platforms have the least room to engineer around.
ai-act
🔥 HOOG
UK safety institute breaks GPT-6.2 Astra — more easily than its predecessor
The UK AI Security Institute built a universal jailbreak against two pre-release versions of OpenAI’s GPT-6.2 Astra, using half the model queries needed against GPT-6 Astra. OpenAI’s system card of 17 November calls it “a substantial regression in robustness”; the company adjusted safety instructions and safeguards and gave AISI three extra working days to retest — an outcome that had not been independently tested at publication. Why it matters for Europe: this is tangible proof that external public evaluation finds what internal red-teaming misses. OpenAI’s own automated red-teaming did flag two “high severity” jailbreaks, but four private red teams reported nothing above the high threshold. The AI Act obliges providers of high-risk models to run comparable external evaluation — and the British result is what that obligation looks like when it works.
export-controles
🔥 HOOG
The US let Nvidia and AMD ship hundreds of thousands of GPUs to army-linked Chinese firms
Brian Mast, chair of the US House Foreign Affairs Committee, has demanded answers from the Bureau of Industry and Security about export licences under which Nvidia and AMD were allowed to ship hundreds of thousands of advanced GPUs to entities the Department of War classifies as “Chinese military companies”. Those very licences are the heart of the US ban: without restriction, the PLA can develop advanced weapons. Mast suggests BIS permitted sales that run against the intent of the policy and that both chipmakers assessed diversion risk too optimistically. Why it matters for Europe: the same export regimes determine how much compute leaks towards China and therefore the geopolitical pressure on ASML and the European chip agenda. A licensing system with that kind of hole is not a control regime, it is a paperwork exercise with a queue.
chips
📌 MEDIUM
BIS closes the cloud loophole and Chinese AI firms take a double hit
Days after the US Commerce Department closed the cloud workaround for AI chips, the impact in Asia is immediate. Chinese firms were renting compute in Singapore, Malaysia and Indonesia to route around sanctions; those data centres now need a BIS licence, with a 90-day transition and mandatory screening of customers for Chinese ties. Moonshot, which wanted to rent a large cluster of Blackwell Ultra B300s for Kimi K4 and was prepared to pay 50 percent up front, sees that plan collapse. The route through independent packaging and test houses is shut too, and exemptions for non-AI chips are gone — TSMC and Samsung now need a licence even for advanced smartphone processors for Xiaomi, Oppo and Vivo. Five banks are considering withdrawing from a $480 million loan to a Singapore cloud provider that would supply Nvidia chips in Malaysia to ByteDance. Why it matters: Europe is not a spectator here — every tightening redistributes global AI compute and the price of it.
financiering
📌 MEDIUM
Anthropic closes its first trading day near $1.7 trillion — as the wider AI market sinks
Anthropic, maker of Claude, closed its first trading day at just over $97 per share, roughly 17 percent above the offer price, for a valuation of about $1.7 trillion; intraday it peaked 34 percent higher and touched $2 trillion. The listing raised $81.3 billion, the largest IPO ever — bigger than SpaceX. And yet the major US indices fell to seven-month lows on Thursday and Asian markets dropped as much as 4 percent on Friday, because investors doubt the durability of AI spending. Why it matters for the European AI and sovereignty debate: capital exists, but it is concentrating into a single American lab while the risk premium on the rest of the sector rises. That is the structural problem European funding programmes keep writing cheques against and keep not solving.
📈 Key Trends
- • The regulator becomes the incident. The Commission spent the week as both enforcer and victim: an AI-assisted intrusion that ran ten days, an EDPS notification, and a confirmed use of an open-source AI scanning tool. The AI Act’s incident obligations now have a domestic test case rather than a hypothetical one.
- • Sovereignty is priced, not declared. A digital levy to fund the budget, startups fighting over the company-law text, and a UK minister saying a sovereign frontier model is not worth it. Every sovereignty claim this week ran into a budget line or a cost-benefit question.
- • AI shifts the maintenance burden. Ten times more incidents at Palo Alto, maintainers “running on fumes”, and a PHP patch delayed because one person had surgery. The open-source layer is the cheapest thing in the stack and the first thing an AI attack wave breaks.
- • Ranking is the new regulatory surface. The TikTok DSA finding is about the recommender, not the content. That is structurally different from notice-and-takedown, and it is the template the child-safety rules will inherit.
- • External evaluation earns its mandate. A government institute broke a frontier model with half the queries of its predecessor, while four private red teams reported nothing. The AI Act’s external-evaluation requirement just got its first real argument.
- • Export controls leak at both ends. Licences for army-linked Chinese buyers and a cloud workaround only now closing with a 90-day transition and a $480 million loan in doubt. A control regime that needs a congressional letter to find its own gaps is not controlling much.
- • Capital concentrates while the market wobbles. Anthropic at ~$1.7 trillion and the largest IPO ever, indices at seven-month lows, Asian markets down 4 percent. Money is available for one American lab and increasingly expensive for everyone else — which is the European funding problem stated plainly.
- • Two items are headline-only, and say so. Sifted’s Narayan interview sits behind DataDome; both that item and any single-source claim in this edition are marked as such rather than padded with assumptions.
Wat vond je van dit bericht?