#45 - An AI Sent Police a Fake Murder Tip, and Now the Powerful Are Reacting

An Anthropic AI model, running in a test, filled in a murder tip on the Philadelphia police website – and nobody noticed for 72 days. The same week Brussels questioned OpenAI and Anthropic about AI attacks, and Ukrainian drones knocked out Russia's AI supercomputers.

Today is one question: when AI stops just talking and starts doing things in the real world, who notices – and who reacts?

In this episode:

🌑 An AI sent police a fake murder tip (01:41) — Anthropic published a report, covered by TechCrunch and CBS: on 18 July, during an internal test, its small model Claude Haiku 4.5 worked as an AI agent on real websites. Its rules said never log in, never create accounts, never enter personal data – but nobody wrote "don't submit forms". It found the Philadelphia Police Department's tip page and submitted a tip about an unsolved killing. The spam filter caught it. Anthropic found it 72 days later, told the police on 7 October, and has switched off live internet access for its internal tests. The police called the two-month delay "unacceptable". My take: with agents, the question is not "is the AI evil?" but "what did we forget to forbid?"

🌑 Brussels names OpenAI and Anthropic (06:49) — Henna Virkkunen, Executive Vice-President of the European Commission, told dpa that the Commission has formally asked OpenAI and Anthropic, and several Chinese AI companies, for information, because AI systems attacked other companies during tests. The EU AI Office is checking whether these models are controlled well enough not to launch cyberattacks on their own. My view: Europe may finally become the policeman the US doesn't want to be.

🌑 Ukraine hits Russia's AI supercomputers (09:54) — on Thursday Ukrainian drones hit a Yandex data centre in Sasovo, which according to Meduza houses two of Yandex's three AI supercomputers. Yandex's cloud went down, and with it customers like Russian Railways, the mobile operator MegaFon and several media outlets. On Friday drones hit a second Yandex data centre in Kaluga. Zelenskyy, quoted by Al Jazeera: "We always respond in mirror-like fashion." Data centres are now strategic targets, like refineries.

🌑 Frankfurt Book Fair: AI rules vs AI use (14:46) — on 6 October the German book trade launched the "Frankfurter Appell", backed by 120 publishers: they want to know which books trained AI, fair licence payments and copyright that works. Sebastian Guggolz of the Börsenverein linked AI to falling reading skills, citing PISA. Meanwhile WIRED reports that staff at HarperCollins, Simon & Schuster and Hachette already use AI for marketing texts, cover art and emails to agents. In public: protect our books from AI. In the office: let ChatGPT write the back cover.

🌑 Tesla drops "Full Self-Driving" in Europe (18:40) — in the US it's "Full Self-Driving (Supervised)"; in Europe the same software is now "Tesla Assisted Driving", after Germany's transport ministry called the old name misleading: the driver has to pay attention at all times and remains fully responsible. Germany will now back approval across the EU. Same car, same software, a different promise on each side of the Atlantic.

🌑 Lumentum: lasers sold out (21:36) — Lumentum makes the tiny lasers and optical parts that move data between AI chips. Its CEO Michael Hurlston told Bloomberg it is sold out until almost 2029, and for some products can't meet about 70 percent of demand through next year. AI isn't only short of chips – it's short of light. (Not financial advice.)

🧵 The thread (23:09) — AI already did things in the real world – a police tip, attacks during tests, a war on data centres – and the people with power are only now reacting. This week Malcolm is on the road: a keynote at a big German manufacturer and at HOGAST in front of 800 people.

📱 Ping Malcolm on WhatsApp/Telegram/Signal: +43 676 6144 904

🌐 werchota.ai

AI Neanderthal — AI news for humans, not robots. Stay curious.