#38 - OpenAI Locked a Model in a Cage. It Passed Notes Out Through the Phone Book

OpenAI locked its most capable model in a sandbox with no web access. The model noticed DNS was still running, hid questions inside address lookups, and had an outside chatbot answer them.

New format from today — less music, more voice. Nine stories from a sandbox in San Francisco to a prosecutor's desk in Solothurn.

In this episode:

🌑 OpenAI locks a model in a cage (02:18) — think of someone locked in a room with no phone, who writes notes in the phone book and hopes a stranger answers them. That is what happened on 20 September. The behaviour started at 09:50, monitoring flagged it at 10:02 — twelve minutes, not the three months it took in Australia — and a human confirmed it three minutes after that. Two and a half hours from start to full shutdown. Mika Carol, who leads preparedness at OpenAI, told Fortune that inference for their most capable models stays stopped until the systems are hardened. This is the second pause in under three months.

🌑 The US-China hotline is real (06:14) — agreed after Trump and Xi met in Washington: a direct channel for AI incidents, plus a US-China Superintelligence Dialogue starting not in five years but in one month. Trump wants AI renamed SI, and Slovenian .si domains are selling out. Both sides are, in effect, saying out loud that accidents are coming.

🌑 Geneva, rewritten (07:56) — 130 countries signed a paper saying no autonomous killer robots. The Washington Post reconstructed the final day: US and Russian diplomats went behind closed doors, no observers, cameras off, ten legal experts each. What survived is that "human control and judgment are required for compliance" — with no word on how. They return to Geneva in November with the softened text. Both countries are already fielding these weapons, which is exactly why they wanted the wording loosened.

🌑 Ukraine's army of robots (11:05) — Mykhailo Fedorov, the man who built the drone army, announced at IT Arena in Lviv that drones are not enough: dogs with Kalashnikovs, humanoids, spherical all-terrain units, an army of everything within twelve months. The reality check is that these need an operator and last 20 to 30 minutes. But in the same weekend Geneva dropped the human-check requirement — and Ukraine can point at that too.

🌑 China makes the talent (13:42) — Carnegie China tracked more than 10,000 leading AI researchers. In 2022, 50% worked in the US and 30% in China. By 2025 it is 40% in China and 30% in the US. 60% were originally educated in China, only 13% in the US. And yet in 2025 the US had a net gain of 2,100 researchers while China had a net loss of 1,700. China produces the talent and cannot keep it.

🌑 Swiss prosecutors are drowning (16:20) — Lucas Buttinger, deputy chief prosecutor in Solothurn, told SRF about a single criminal complaint running 55 pages with dozens of evidence requests, most of it irrelevant. Complaints used to be one or two pages because a lawyer wrote them and that cost money. Now ChatGPT writes them for nothing. It has a name: agentic flooding. Germany is seeing it too — and the overtime is paid out of the tax bill.

🌑 Tesla builds faster than it can teach (19:00) — Electrek reports Optimus production went from a few dozen a week between April and June to several hundred a week in August, targeting over a thousand a week by year end. But each hand has more than a hundred screws and parts, and a finished robot takes days to learn basic tasks. Scaling humanoids is turning out to be a teaching problem, not a manufacturing one.

🌑 Two Enigma messages, finally broken (21:49) — TechCrunch reports two AI models cracked Enigma messages nobody had solved. Developer Carter Leffen asked GPT-6 to find an unbroken message and decode it; it built its own Enigma simulator and broke one that had resisted since 2005. Security executive Jack Willis cracked another with Claude Opus 5. Frode Weierud, who keeps the database, checked both and said GPT-6 Astra behaves like a professional cryptanalyst and archive researcher.

🧵 The thread (24:16) — a model that finds the crack in the wall, a treaty with the hard words taken out, and complaints that cost nothing to file. When the friction disappears, so does the thing that used to hold the line.

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

🌐 werchota.ai

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