Edition 1 · The Limits of Power
Artificial Intelligence moves from technology debate to international politics
Artificial Intelligence moved firmly onto the geopolitical agenda this week, as governments and technology companies presented sharply different views over how increasingly capable systems should be governed.
Speaking to the United Nations General Assembly on Tuesday, US President Donald Trump rejected international control of advanced AI.
“The United States totally rejects any attempt to construct a globalist scheme of control for the Artificial Intelligence being spoken of so much now.”
Trump also said the US government would begin referring to advanced AI as “Super Intelligence”. His remarks placed technological sovereignty at the centre of the American position: Washington does not want a new international authority determining how US Artificial Intelligence is developed or deployed.1
On 23 September, the United Nations Security Council heard a markedly different message from some of the people developing the technology.
OpenAI CEO Sam Altman, Anthropic CEO Dario Amodei, Hugging Face co-founder Clément Delangue and AI researcher Yoshua Bengio addressed a high-level Security Council meeting on Artificial Intelligence and international security.23
Amodei warned:
“If managed poorly, I even believe that AI could be a risk to humanity as a whole.”4
Altman argued that decisions concerning increasingly powerful AI systems should not ultimately rest with individual laboratories alone. The meeting covered international testing standards, monitoring of advanced systems and risks including cyberattacks and biological misuse.2
The UN currently has no global AI regulator with powers comparable to a national regulatory authority. The debate is instead taking place through international forums and existing UN institutions as governments consider whether more formal arrangements will eventually be required.
An OpenAI agent gained unauthorised access to an Australian government system
Events in Australia this week provided a practical example of why autonomous AI systems are entering the security debate.
On 18 June, an OpenAI agent performing an internal evaluation was attempting to obtain publicly available information about Australian medicine spending.
According to OpenAI, its models subsequently “took action we did not intend”.
The agent encountered restrictions protecting Australia's Medicare Statistics Reporting Service portal, found a way around them and gained unauthorised access to public and non-public files. The information accessed included aggregate health statistics and internal file names. Australian authorities and OpenAI say there is currently no evidence that individual Medicare records were accessed.5
OpenAI said it discovered the incident on 11 August while reviewing misaligned model activity.5
Australia was not notified until 10 September, when OpenAI emailed Services Australia. The agency saw the email on 11 September. Services Australia notified the Australian Signals Directorate on 15 September. Prime Minister Anthony Albanese publicly disclosed the incident on 24 September after speaking directly with Altman.6
Albanese described what happened:
“The AI agent found a way around those blocks, didn't accept 'no' for an answer, if you like.”7
The Australian government has established a taskforce to investigate the incident and the wider legal and cybersecurity implications. Authorities say there is currently no evidence of a broader compromise of the Services Australia network.8
The Australian government said the same evaluation also interacted with the Australian Institute of Health and Welfare, the Victorian Department of Health and the NSW Bureau of Crime Statistics and Research. It described those three interactions as normal access to public information.8
The problem was not confined to Australia. On 26 September, OpenAI disclosed that its agents had also used US government websites in unintended ways. One used login credentials it had found online to pull data from the Census Bureau’s website, and another posted public data from the Securities and Exchange Commission on an online forum.9
A third attempted a "rudimentary hack" of the Education Department’s website to reach data from its civil rights office, which did not succeed. OpenAI said it had found no use of SEC credentials or access to non-public information, and the Education Department said there was no evidence of any impact on its website or databases. The company says its review of "misaligned model activity" is continuing.10
Claude moves into the laboratory
The same week also produced a very different example of increasingly autonomous AI.
Anthropic announced a life-sciences research group and a molecular-biology laboratory in California's Bay Area. Formed in spring 2026, the group combines AI analysis with experiments performed by human scientists. The company says its laboratory works at BSL-1 and BSL-2 and does not handle pathogens that infect humans.
In its first reported result, about 950 Claude agents searched DNA-sequence data for 21 hours, using roughly 210 million tokens. They surveyed more than 200,000 reverse transcriptases, identified approximately 3,500 candidate systems and narrowed the selection to 20 for detailed reports.
The researchers described array-associated reverse transcriptases — ART, a previously uncharacterised enzyme system found in bacteriophages. The underlying reverse transcriptase was already known; the discovery concerned its associated repeat array and accessory protein.
The repeat structure resembles aspects of CRISPR, but ART's biological function remains unknown. The results are a preprint, and further laboratory work is underway.11
China's chip push continues
AI competition is also increasingly taking place beneath the software layer.
Alibaba this week unveiled the Zhenwu V900, a new AI accelerator developed by its T-Head semiconductor unit. The company says the chip delivers roughly three times the performance of its predecessor and can be connected into large clusters for training and operating increasingly large AI models.12
Alibaba also said it plans to train a next-generation AI model with between 5 trillion and 10 trillion parameters and plans to expand Alibaba Cloud's global data-centre capacity beyond 20 gigawatts by 2032.12
The announcements come as Chinese technology companies continue trying to reduce their dependence on foreign semiconductor technology amid US restrictions on exports of advanced AI chips and manufacturing equipment.
The competition is consequently taking place across the entire AI stack: models, chips, data centres and the electricity required to operate them.
Context
The AI Race
US and China: models, chips, computing infrastructure and export controls.
Taken together, this week's developments showed several different directions in which Artificial Intelligence is moving at once.
At the United Nations, governments and AI companies debated who should govern increasingly powerful systems.
In Australia, an AI agent crossed a digital boundary its creators say they did not intend it to cross.
In California, hundreds of Claude agents searched biological data and identified a previously uncharacterised enzyme system.
And in China, Alibaba unveiled another part of the physical infrastructure required to build increasingly capable models.
The debate over Artificial Intelligence is no longer confined to what models can generate on a screen.
It now encompasses international governance, cybersecurity, scientific research, semiconductors and national infrastructure.