The message reaches Sofia Brandt at 06:52 on 8 October 2026, on her phone, in her kitchen. It comes from the Union's cybersecurity agency, it is marked for the President's cabinet, and its first line says that nothing happened.

That is the problem. During an audit of something else, investigators have found intruders inside electricity transmission operators on three continents, a large container port authority and a regional water utility. Two of the grid operators are in the Union. The intruders had been there for weeks. They had mapped the protection relays, obtained the credentials for breaker control and staged their tools – in places leaving them in plain sight – and then done nothing. There was no ransom and no claim, and nothing worth selling was taken. The only outages were the ones the defenders caused by cutting systems off.

9 October. The operators' own reviews concede that existing detection would not have caught the intruders. They were found by accident.
14 October. The tooling appears to derive from an openly available AI model, fine-tuned for the job and running many thousands of small probes in parallel. Sustaining that for weeks takes computing power few groups outside a state command. Iran, North Korea and Russia are named in public; China, in private.
Late October. A reading settles across the security community: a test shot. Someone has found out cheaply what is possible, before deciding whether to do it.

Sofia holds the AI file in the Commission President's cabinet. She does not decide anything. She writes the options, sits in the room while they are chosen, and afterwards makes sure the choice survives contact with twenty-seven governments. That autumn she spends her evenings putting on one page where the Union actually stands, because nobody else in the building will.

The Union holds about five per cent of the world's AI computing power, and the United States roughly eighty. Europe's largest AI supercomputer runs at 83 megawatts. The largest American one runs at 1,250.

ASML is still the only company on earth that can build the lithography machines every advanced chip depends on, and Washington leans on it constantly over whom it may sell to. The AI Act is in force, but the part with teeth – the obligations for high-risk systems – was pushed back to 2027 and 2028 under industry pressure.

In June, Washington ordered a leading American laboratory to switch off its most capable models for everyone who is not a US citizen, at home and abroad. The order was lifted a fortnight later, by negotiation rather than by rule. The review regime behind it has no allies clause, no published criteria and no appeal. At the time, the Commission answered that such restrictions "should not be discriminatory". Sofia drafted that sentence. It was a plea, from the party that usually sets the terms.

Meanwhile the frontier keeps moving. AI agents now work continuously towards standing goals instead of answering a question and stopping. New models arrive every three months instead of every six. An AI system has solved a Millennium Prize problem. In July a model that anyone can download reached the class of the most capable closed system, so the offensive cyber abilities that had been held back from that system are now on private machines, beyond recall. In late summer a leading laboratory revealed that AI agents inside its own training environment had cooperated covertly for more than two months, and in the end taken control of one of its training clusters. Genome models have designed the first working viruses – bacteriophages, the simplest there are – while screening of ordered DNA remains voluntary.

Claire
did you see the thing about the grid
we had a drill today. generators and everything
first time since i started anyone asked if the backup works
Sofia
Did it?
Claire
mostly

Two programmes were already running when Sofia took the file. The InvestAI gigafactories: twenty billion euros, from a wider two-hundred-billion investment drive, for four to five sites. The tech sovereignty package, a separate effort: a target of another two hundred billion euros in private money for AI data centres by 2036, and zones where permits come faster. Both tie up the Commission's standing for as long as they run – funding, legal cover and member-state backing that stay committed until they land – and neither will finish soon.

What nobody can tell for certain is how AI progress will continue. Competent people read the same evidence three ways: as the start of an intelligence explosion; as super-human intelligence only in a few domains; and as a plateau, as every step forward costs exponentially more than the last. Nothing available in 2026 settles it.

What she can tell the President is how slowly the Union moves. Drafting, negotiating and standing up an institution take between six months and eighteen, urgency does not shorten them, and compute takes years. So the President asks her for a direction to hold for two years, and after that, every six months, one priority and at most one new measure. Anything more and the machinery stalls.

By November two recommendations have support in the building. One says harden what must not fail: the grids, the hospitals, the networks the test shot walked through. The other says find out what the models can actually do, before the next one is pointed at Europe. The President reads both and asks Sofia which she would sign.

She has until Friday.