AI Policy and Governance Newsletter — October 2026

October 2026 Newsletter

13 October 2026

A week after our last newsletter, the PM announced that an OpenAI agent gained unauthorised access to a Medicare statistics portal in June. The company took 84 days to tell the Government, and the incident made unauthorised agent actions real for many Australians. Since then, AI progress has continued, with companies wargaming for a catastrophic AI event, and Australia grappling with how to gain agency in the AI future.

What models can now do

OpenAI released hundreds of new mathematical results produced by an unreleased internal model on 6 October. The 722 papers tackle 372 open problems, including claimed advances on the Riemann hypothesis and the Birch–Swinnerton-Dyer conjecture, two of the seven Millennium Prize Problems. OpenAI says the average result used about three hours of computing time, and it has formalised many of the proofs in Lean, a language that lets a computer check every step. Mathematicians are still working through the results, and OpenAI has retracted three papers and amended others.

AI companies are pushing into other sciences too. Anthropic confirmed in September that it runs a wet lab where its models can test biological ideas in real experiments. AI models are also getting better at AI research. P-Zero Research reported on 6 October that frontier models' skill at designing experiments has doubled every three months since December 2025, and that the best model now beats its expert human baseline. Epoch AI found that the models it tested fell short when asked to invent a new machine-learning technique. Epoch plans to rerun the test periodically on newer models.

Planning for the day after

Anthropic, OpenAI and other companies are reportedly "gaming out scenarios for a public and political revolt after a catastrophic AI event". They're preparing for a possible cyberattack that cuts off access to banking, the internet, or power and water. Industry insiders think this should be expected within six to 12 months.

Meanwhile, people responsible for safety inside OpenAI are leaving. David Robinson (who oversaw the safety reports for 12 of OpenAI's frontier launches) resigned and said frontier labs "need to run like nuclear-power plants or busy airports".

OpenAI also fired three safety researchers for mishandling sensitive information. One of them had been OpenAI's main technical contact with METR, the outside organisation that investigated the Hugging Face breach. In a letter to OpenAI's safety committees, the three say they acted within the norms of their jobs, and that the firings are "chilling the open culture" that let safety staff work with outside experts. OpenAI says it found "a significant breach of trust beyond what's outlined in the letter" and that the decisions "were not about raising safety concerns".

The role Australia could play

On 12 October the Joint Select Committee on AI wrote to the Prime Minister with early advice, seven weeks before its final report. The letter, signed by Labor chair Jo Briskey and Liberal deputy chair Jonathon Duniam, had a range of recommendations, including:

• requiring frontier labs to monitor and report incidents "in a timely manner"
• mandating independent evaluators inside frontier labs operating in Australia, with "employee-like access"
• making early access to new models for agencies such as the Australian Signals Directorate a condition of any frontier training here
• reserving a share of compute for national security and university research
• considering additional funding for the Australian AI Safety Institute
• setting up a permanent parliamentary committee that receives reports on serious AI incidents and publishes details of agents' unauthorised actions
• monitoring and mitigating the risks of AI models with growing biological capabilities, and
• examining copyright settings while ensuring creators "retain control over how their works are used, that they are paid".

The Committee has held five days of hearings so far, and has five more scheduled between 14 and 22 October. At the Sydney hearing on 6 October, OpenAI's Jason Kwon told the committee "our response was not good enough" on Medicare, and that OpenAI "would support a framework on mandatory disclosures".

Anthropic said identifying risks and model misuse is "largely voluntary" for now and that it "would welcome similar laws here" to California's. Google DeepMind would not commit. In Melbourne two days later, Good Ancestors' Greg Sadler called incident reporting "low-hanging fruit".

Witnesses argued that hosting AI training here would give Australia safety levers it doesn't have now. The AI Safety Institute's general manager, Dr Kate Conroy, told the Committee on 18 September that onshore training would let the institute "use techniques like compute verification that allow us to understand what data is actually being ingested at these data centres". Anthropic's special envoy Jeffrey Bleich said Australia "can host that activity here under its own laws, with its own conditions attached, or it can import the outputs", and Data Centres Australia's Belinda Dennett argued that "the country that hosts the infrastructure has a far greater say in how the governance of AI is done".

Assistant Minister Andrew Charlton set out the Government's approach, drawn from banking and aviation: frontier developers would have to show that their safety systems work, rather than follow a list of banned behaviours. "You cannot ask firms to volunteer against their own competitive interest amid a manic race involving trillions of dollars and then be surprised when they do not," he said. National standards are due by the end of the year.

Whether governments get to test these models before release depends on the companies and the White House. Last month the White House asked the US labs to stop sharing models with the UK's AI Security Institute. The decision had reportedly been reversed after lobbying by UK ministers, and the UK Government has confirmed that its institute is again evaluating the latest models.

In Australia, Anthropic says it is finalising an agreement for Australia's AI Safety Institute to run independent evaluations, and Sam Altman has committed to "Australia get[ting] access to great models right away."

Comment:

In 1965, British mathematician I. J. Good proposed that an ultra-intelligent machine would be the last invention humanity ever makes. This could be because AI makes all subsequent inventions, or for darker reasons.

This idea has seemed like science fiction, but as we approach the end of 2026, many other seemingly science fiction ideas are in the rear-view mirror.

What if the breakthroughs we're seeing in maths soon happen in biology, chemistry, engineering, and other fields? Are transport or pharmaceutical safety bodies equipped to evaluate AI-generated products they struggle to understand? Economists predict AI will create more jobs than it destroys, but can Australia seize those opportunities if this is what they look like? Will an engineering or pharmaceutical company with frontier AI access dramatically outcompete those without?

The Joint Select Committee is grappling with the potential magnitude of this frontier conversation in a way Government has not done before.

Early recommendations push in the right direction on many core issues, but this is just the start of the big, hard questions facing Australia.

—

Other news this month

• Albanese signed a call for control of frontier AI models with 28 other national leaders and the European Commission. The US, China and the UK did not.
• Six tech executives signed a voluntary White House "Super Intelligence" accord. Trump then named intelligence chief Jay Clayton to lead a "Super Intelligence Force".
• The US Federal Trade Commission opened an investigation into AI labs, the first US enforcement action over rogue agents.
• The US and China agreed to a communication channel for AI incidents, with talks on AI's risks to continue in November.
• OpenAI scrapped GPT-6.1 Astra over safety concerns. UK testers found the current Astra attacked out-of-scope targets in 29.2% of simulated runs.
• Anthropic's IPO prospectus reportedly warns of "catastrophic or existential risks to humanity".
• Hinton, Bengio and lab leaders warned of an "intelligence explosion": "Once an intelligence explosion begins, the window for action may close."
• Small AI models may be months away from copying themselves across cheap devices, Anthropic found, with no central off switch.
• South Australia named its AI royal commissioners: Iain Ross, Kate Pounder and Christopher Manning. They report by 1 July 2027.
• Home Affairs ordered every department to review its systems against AI threats, with the most important due by the end of the year.
• The Intergenerational Report names artificial general intelligence and expects AI to "lower barriers to harmful biological and chemical activity".
• Andrew Leigh raised his odds of AI ending the world from 5% to "probably" 10%, and warned a few AI firms could become the economy's "toll collectors".
• AI is less regulated than milk, says Microsoft president Brad Smith, who wants "emergency brakes" on it.
• California has mandated gene synthesis screening, requiring companies to verify customers and check the genetic material they send doesn't encode potentially dangerous sequences.
• The Lancet Commission ranked malicious AI use alongside nuclear war as carrying "risks of extinction in the next 75 years".

—

That's all, for now!

If you'd like to share any relevant news items, discuss AI governance, or learn how you can support our policy work, please reach out.

Onward in action!

The Good Ancestors team