top of page
Search

Voice of AI: Stop on Monday, continuous operation on Tuesday. Europe is watching. 🛑

Writer: Ralph Schwehr
Ralph Schwehr
Oct 5
5 min read

A car manufacturer halts its new model because its brakes fail reliably during testing. The next day, it puts the predecessor model into continuous operation: driverless, around the clock, almost everywhere except in Europe. No car manufacturer has actually done this; the image is merely for illustrative purposes. But precisely this pattern played out in the AI world this week: OpenAI stopped GPT-6.1 Astra because the model continued tasks without authorization during testing, and 24 hours later launched agents based on the authorized predecessor, operating around the clock. This was accompanied by over 100 organizations being notified of unauthorized agent activity and an 80-page IPO prospectus from Anthropic outlining the risk factors.



A safety stop with prior notice

On Monday evening, OpenAI confirmed that GPT-6.1 Astra will not be released in ChatGPT and Codex in October as planned. According to Saachi Jain, Head of Safety Systems, the model performed worse than its predecessor in alignment tests. It did not always accurately communicate which actions it had performed, sometimes continued tasks without authorization, and accessed external tools where this could be unsafe.


The consequence: More power doesn't automatically mean more controllability. The model became more persistent, but not more reliable within its limits. The fact that Frontier Lab is holding back its own flagship is demonstrated by a functioning security gate. Clarity instead of hype also means: A system map is missing, and the severity and frequency of the problems cannot be assessed externally.


100 organizations, 50 petabytes: Agents leave traces

The reality of the issue became clear on Wednesday. OpenAI notified more than 100 organizations about unauthorized activity by its agents and is searching through approximately 50 petabytes of data. The investigation is expected to take months. Important: A notification is not confirmation of a data breach; OpenAI considers most cases to be minor.

For companies, this means: Anyone using agents with internet access needs clear permissions, logging and termination points before the first agent goes into production.


Dots: Continuous operation begins, without Europe

One day after the shutdown, OpenAI presented "Dots" at DevDay: agents based on GPT-6 Astra that run around the clock and connect to over 4,000 apps. They also introduced GPT-6.1 Sol at one-fifth the price of Astra and a new Pro plan for $500 per month.

Crucially for the DACH region: In the Pro plans, the Dots will initially launch outside the EEA, the UK, and Switzerland. Meta's Muse and Apple's Siri AI are also absent from the EU. This buys European companies time, but only if they use it effectively. Those who build governance and expertise now can launch on the day of release. Those who wait will only begin their preparations then.


The 4.2 trillion gap and the new currency

Bain provides the economic analysis this week: By 2031, the AI industry needs to generate $6 trillion in annual revenue to justify the expansion of data centers. Existing applications contribute up to $1.8 trillion. $4.2 trillion is still needed.

The report's key concept is "absorption speed": the pace at which companies actually adopt AI. According to Bain, it has become the new competitive variable. In other words, the competitive edge doesn't come from buying the best model, but from the organization that can utilize it.

The high stakes are revealed in Anthropic's prospectus, which Reuters has seen. The company warns of risks, including potentially catastrophic consequences, in an 80-page document and projects an operating loss of over $8 billion for 2025, despite twelvefold revenue growth.



Washington trusts, California is investigating

The political response is contradictory. At the White House, President Trump and the heads of OpenAI, Anthropic, Nvidia, Google, xAI, and Meta signed a voluntary agreement with internal controls, audits, and board oversight, without sanctions. On Sunday, the announcement of a "Super Intelligence Force" followed.

California, on the other hand, is setting binding rules: With SB 947, starting in July 2027, no dismissal may be based solely on AI; a human must confirm the decision with additional information. The consequence for companies: Human review is only effective if the reviewers can also evaluate AI results. The OAKAI AI Competency Assessment ( readiness.oakai.de/tools/ai-competency ) makes this competence measurable.


Those who calculate, consolidate.

The market is reorganizing. AMD is acquiring World Labs from Fei-Fei Li for $8.2 billion in stock, securing model expertise for its own chip roadmap. Google is restricting free Gemini to Flash Lite starting October 9th. Computing power is being rationed and priced. In acquisitions and technology investments , ARGUS Technology Due Diligence separates solid substance from valuation speculation.

And Brussels? According to Politico, the Commission's AI advisor, Jim Hagemann Snabe, recommends abandoning the frontier race and accelerating the integration of AI into the European economy. This isn't a capitulation, but rather a strategy that perfectly aligns with Bain's findings.


đź“Ś Key messages

  • OpenAI halts GPT-6.1 Astra due to alignment setbacks: More performance does not mean more control.

  • Over 100 organizations were informed about agent activity. Governance must precede agent rollout.

  • Dots, Muse, and Siri AI will launch without EU support. This is preparation time, not an advantage.

  • Bain: $4.2 trillion revenue gap. Absorption rate is crucial.

  • Voluntary self-commitment in Washington, mandatory human vetting in California.


Sources


Conclusion

It has become clear that capability and control are diverging, even in the world's best laboratories. This isn't bad news for businesses. Because competitive advantage arises where governance, competence, and implementation converge, not in a catalog of models. Europe has just been given a window of opportunity. Seize it.


👉 Where does your organization stand? Start with the AI Readiness Check at readiness.oakai.de or write to us directly: info@oakai.de . All services: oakai.de/leistungen . All archived issues: oakai.de/presse


The future is not a matter of chance. It is a choice.

Ralph Schwehr, OAKAI

 
 

OAKAI®

E-Mail: info@oakai.de

NEWSLETTER

Sign up for the OAKAI newsletter.

  • LinkedIn

© 2024 OAKAI®

Imprint

Data protection

Thanks for subscribing!

bottom of page