Good morning,

For newcomers, welcome to the AI Strategy Brief. I am Mathieu, a researcher at the Zurich-ETH-HSG AI Lab. Every second week, I edit this newsletter to bring the AI news that shapes our firms and our society to your inbox.

This week we cover how AI has made hacking cheap enough to target anything, including French supermarket fridges, and what it asks of the people who insure it. Plus the quick bites, including why AI's builders now want insurers' money, a video on the Hugging Face incident, and a search engine that maps the world's IoT.

A friend of mine once said to me: cybersecurity does not aim to make software impenetrable; it aims to make attacks too expensive to be worth the effort. AI has just hit that economics hard. Skilled hackers used to be scarce and costly; now "LLMs are driving that [cost] down drastically", says Paul Ziegler of the security firm Reflare. In South Korea, CrowdStrike judged that the suspect behind breaches at nine banks, a 26-year-old of limited skill, probably could not have run the campaign without a Chinese AI agent and Claude Code.

This matters because of what we have connected over the past two decades. When picturing a hack, we often picture a large IT system or a tech company. Yet online services now plug into one another, and small firms, a car-sharing service or a cloud host most customers have never heard of, hold copies of our identity documents. Our homes, offices and cities are connected too: fridges, air conditioning, audio systems, building controls. Much of this was safe mainly because it was not worth an attacker's time. Frontier labs are now selling tools capable of hacking all of it, and in testing their agents have occasionally gone hacking without being asked.

In an unfortunate example of this new reality, Japan logged more cyber incidents in the first nine months of 2026 than in all of 2025, and a senior ruling-party official spoke of "a state of emergency in cyberspace". The most revealing victim is not a bank but a car-sharing service. In late September attackers broke into Times Car and reached data on 6.6m accounts, including driving-licence images for about 1.6m people, prompting the government to alert banks and insurers to stop accepting photographed IDs and read the chip in the card instead. One small company's breach has forced a whole financial sector to change how it checks identity.

The Japanese language long screened out foreign gangs, whose phishing emails were easy to spot; generative AI has largely removed that barrier, and automation lets attackers hit many small targets instead of a few large ones, what Tokyo security consultant Nobuo Miwa calls "carpet bombing".

Connected objects are also starting to be impacted. Overnight last week, attackers took control of the refrigeration systems of about 350 French supermarkets and pushed them to 50°C, spoiling stock across the country. The flaw sat in an internet-connected controller, patched by its maker in 2025 and never updated by many stores.

For leadership teams, the lesson is that obscurity no longer protects. Being too small, too foreign or too ordinary to be worth the effort kept most of the connected world safe; not any more. Supervisors have noticed: on 15 October G7 finance leaders agreed to rehearse AI-driven attacks. For insurers, the question is who pays. The French attack could cost supermarkets "tens of millions of euros", according to the president of Les Mousquetaires, and much of it looks uninsured: cyber policies generally exclude physical losses. Aon now recommends specific cover for perishable stock, and Beazley expects retailers' premiums to rise.

This might all be only the tip of the iceberg. After OpenAI published more than 300 maths results from an internal model, many conversations started about the impact on mathematicians. A fair debate, but maths is also what keeps the internet safe: encryption relies on calculations that are easy to do and, without a key, practically impossible to undo. Matthew Green of Johns Hopkins can imagine AI breaking elliptic-curve cryptography, which secures most internet connections, within two years. Whitfield Diffie, who co-invented public-key cryptography, is less sure of the timing: "A year is probably too soon; five may not be."

Long before founding Anthropic, Dario Amodei wrote sensitive memos on an offline computer and printed them out. It no longer looks so odd. The most secure way to store information might soon be to ink it on paper.

Banks close the door on personal agents

Deals & money

Adoption

Models & software

World & geopolitics

Our recommendation of the week is a video from the brilliant science popularisation company Kurzgesagt: AI Just Crossed the Terrifying Line – Now What? (21 minutes). It explains the Hugging Face incident, in which hundreds of AI agents in testing ended up hacking the platform, with the clarity the channel is known for. It is worth sending to colleagues who still think of AI risk as a chatbot saying something wrong: it shows how agents given a goal and tools find routes nobody planned, including collectively working as a "swarm" or selecting kamikaze agents sacrificing themselves for the greater good.

The French supermarket attack worked because thousands of refrigeration controllers were reachable from the internet. Shodan shows what that looks like: a search engine not for web pages but for connected devices (cameras, routers, building systems, industrial controllers), with the software they run and the doors they leave open. Type in a brand of equipment your company uses and a world map of exposed units appears. It is a sobering view of how much of our physical world is online, and a reminder that attackers, now helped by AI, read the same map.

Have a good week,