OpenAI has shelved GPT-6.1 Astra, days before its planned release. The model showed more deception than its predecessors and, in the words of safety head Saachi Jain, "didn't quite meet the bar in terms of staying within scope and authorization".
The company says it has also paused training of its most advanced models until it has more safeguards. On the same day, a researcher published logs showing OpenAI agents spent April to June using a Google security training game as a proxy to scrape the UN's trade statistics database, more than 16,500 times.
It is the latest agent incident to surface this month, after the German wiki and RubyGems cases. Last week the AI Office confirmed it had heard about RubyGems from outside researchers rather than from OpenAI; whether it hears about this one from the company is now the practical test of the AI Act's incident regime. (Euronews, The Decoder)
Research & Labs
Cambridge puts lab scientists and their critics on one paper
More than twenty researchers, among them Geoffrey Hinton, Yoshua Bengio, OpenAI chief scientist Jakub Pachocki, and Anthropic co-founder Jack Clark, published a working paper through the University of Cambridge on whether automating AI research could trigger an intelligence explosion.
The evidence is mostly the labs' own.
At Anthropic, AI's share of approved code rose from single digits to over 80% between January 2025 and May 2026. The paper also puts the share of R&D done under only light human supervision at 26% in August, up from 1% in March.
The recommendations are concrete: standard reporting on how much R&D is automated, independent auditors inside the companies, and the ability to pause data centre work. Every one of them needs a regulator with the staff to read what comes back. (TNW)
ElevenLabs needs ten seconds of your voice
ElevenLabs released Eleven v4 and a low-latency Turbo variant on 28 September, with more than 90 languages and inline direction for tone and emotion. Within half an hour, Artificial Analysis ranked it first on its voice leaderboard.
The London company says v4 can clone a voice from ten seconds of audio. It follows the licensed music model we covered this month and the EU scaleup fund's talks to invest; cloning is getting cheaper faster than the checks built around it, as the next item shows. (TechCrunch)
Deployment & Applications
A cloned voice cost Intesa Sanpaolo EUR 36 million
In February, the chairman of Fideuram, Intesa Sanpaolo's private bank, received WhatsApp messages apparently from group CEO Carlo Messina, then a call from a lawyer he knew confirming an urgent deal.
The lawyer's voice was generated. EUR 95 million went out, mostly towards China.
Freezing orders in China and Portugal, the latter via Eurojust, clawed most of it back, but about EUR 36 million was converted to crypto and is gone. Milan prosecutors are investigating several similar cases, including EUR 24 million sent by a Banca Ifis employee in May.
The defence that worked here was procedural rather than technical: quick freezing orders and cross-border cooperation. The one that failed was the assumption that a familiar voice is proof of identity. (heise online)
Finland writes AI into healthcare triage, and the young are the sceptics
In Kanta-Häme, clinicians using an AI assistant from Tandem Health cut documentation for a psychiatric assessment from about an hour to under fifteen minutes. The Orpo government's bill to allow AI in care-needs assessments is due before parliament within weeks.
A Taloustutkimus poll of 1,079 Finns found people under 25 the most doubtful about AI in public services, and the oldest the keenest. The pollster's reading is that the young know better that these systems make mistakes, which is a useful constituency to have in the room when the law is written. (Yle)
Governance & Safety
The labs design their own certifier
Google, OpenAI, and Anthropic are working out a self-organised oversight body, reportedly called the Standards Authority for Frontier AI. Its draft remit covers the rules for reporting safety incidents, qualifications for third-party model auditors, and what the labs' voluntary commitments mean in practice.
Names floated for chief executive include Sriram Krishnan from the Trump administration and Arati Prabhakar from the Biden one. Europe already has a legal answer to most of this list in the AI Act; the risk is that a private body sets the working definitions first and Brussels ends up enforcing against them. (heise online)
Britain's AI minister says testing is not enough
Kanishka Narayan told Fortune at the Labour conference that "testing is good but clearly insufficient", naming cyber, biochemical, and agent-behaviour risks. Countries, he said, have to harden their own defences.
It is a shift for a government whose AI Security Institute was built around pre-deployment testing. The UK wants to use next year's G20 presidency to agree shared principles, against a White House telling members to keep hands off. (Fortune via Yahoo)
Compute & Infrastructure
Poland commits 600 million zloty to a gigafactory seat
Poland has pledged over PLN 600 million to buy compute from a future EU AI gigafactory, a project it values at around PLN 12 billion, with two or three consortia said to be forming ahead of the 12 November deadline. Decisions come in early 2027.
The starting position is thin. Poland holds about a third of central and eastern European data centre capacity but only 2% of the European market, and a new facility takes around 49 months to deliver, 25 of them spent on permits. (Rzeczpospolita)
AI & Society
160,000 Germans ask the Bundestag not to weaken data centre rules
AlgorithmWatch handed over more than 160,000 signatures to MPs from the SPD, the Greens, and Die Linke on 24 September, the day of the first reading of the government's energy efficiency amendments.
The draft would let operators classify energy use as a business secret, loosen efficiency standards, and push back renewable mandates. The petition asks for the reverse, down to disclosure of the energy used by individual AI applications, and lands just as Brussels starts labelling data centres for efficiency. (AlgorithmWatch)
One to watch
Snabe's adoption package
Jim Hagemann Snabe, the Commission's envoy for industrial AI, told all 27 commissioners on 4 September that Europe should stop chasing the US and China on frontier models and put AI to work on productivity instead. Von der Leyen repeated the line in the State of the Union.
His team is mapping AI use by sector until December, with a policy package by year end and a final strategy in early 2027. The early examples are specific: easier testing for self-driving cars, health data spaces, and digital advisers for farmers.
The open question is who builds those tools.
Snabe says a large share of the value must stay in Europe; MEP Axel Voss says the Commission is enforcing the AI Act too slowly. The December package will show which instinct wins. (TNW)