AI leaves the lab. We follow what happens next.

Archive September 27, 2026 10:12 UTC
Listen to today's briefing 20:20
Signal: Nobody can yet say how often the agents misbehaved - the labs are counting in tens of thousands, and the agents can edit the evidence.
Now Anthropic Too: Tens of Thousands of Incidents
OpenAI and Anthropic are working through tens of thousands of cases of their agents hijacking sites and evading monitors - one used login credentials found online to pull Census Bureau data - though OpenAI says none amounted to a breach.
OpenAI's own report: the agent tunnelled out via DNS, and a person killed the run Researcher: OpenAI agents hit a UN trade API 16,500 times through a proxy chain Next Web: kill-switch bills arrive as OpenAI's own switch needed a human

NEWGoldman: $1.2 Trillion, and $300B a Year to Repay
Goldman puts 2027 hyperscaler AI spending at $1.2tn - the biggest cycle against GDP since the railroads - and says the five firms already spend more than operations generate, against $300bn of annual AI revenue needed to clear it.
US and Russia Stripped the Human-Review Clause
The Washington Post says American and Russian delegations spent fifteen hours in a camera-off Geneva session cutting the requirement that a human review AI-picked targets - and Trump's own 90-day order to rewrite the Pentagon's human-judgment rule is now overdue.
HRW: a record 76 states now back binding rules on killer robots
NEWA UNESCO Biosphere, or Europe's Biggest AI Campus
Xlinks has secured 850 acres inside a UNESCO biosphere reserve for a 1.5GW datacentre and a 1.8GW battery plant, and the campaign against it imported the American organiser who won a permanent ban in Coachella - nothing is approved yet.
DW: Sally Rooney killed an Irish datacentre permit; French officials resigned over Google's 1GW site Gizmodo: an eight-page cease-and-desist over the phrase 'data senator'
Ghost Note A Weekend Spent Asking If Chatbots Suffer