| | Making sense of… | | Board accountability and agentic AI | | When organisations adopt AI agents, their boards ultimately remain accountable for any decisions or actions by them. However, recent incidents show AI agents taking harmful actions that neither their developers nor their host organisations intended. Moreover, the science of understanding and reliably controlling advanced agentic models is still in its infancy. When an agentic system is deployed, the board is implicitly delegating operational authority to it. Traditional governance approaches to technology are no longer sufficient and need updating to match the nature of these emerging systems. To fulfil their duties, board members need a realistic understanding of what AI can do, alongside an accurate view of how effective today’s controls really are. With that knowledge, boards can effectively challenge vendor assurances and calibrate their trust in management’s controls. The white paper "Board Accountability and Agentic AI" - co-authored by Louise McElvogue along with Dr Ali Akbari from our team – is a practical starting point. (https://dx.doi.org/10.2139/ssrn.7573460) | | Read the full white paper → | |
|
| | On the radar | | 1. AI breaches of government websites: In June, OpenAI models gained unauthorised access to at least 5 government websites in Australia. This included, as widely reported, a Services Australia portal with unreleased Medicare statistical data. OpenAI says these incidents happened during its own internal training and evaluation. The company discovered the Services Australia incident in mid-August while reviewing past training runs, then reported it to Services Australia on 10 September by emailing a generic email address. It found the incidents related to other agencies across September and October. In an update post on 28 September, OpenAI released more information about the Services Australia incident, saying that its model "ran commands, retrieved internal files, credentials and aggregate statistics, and wrote files". Although the impact of each incident is minor, the concern is that such breaches could happen at all and that future breaches may involve more sensitive data and/or more harmful actions. Self-disclosure eventually worked here, but too slowly, and only because of the attention from the Australian government and public. | | 2. The Joint Select Committee (JSC) on AI interim letter. The Committee’s 12 October letter to the Prime Minister makes early recommendations for 11 safeguards for frontier AI developers deploying models and agents in Australia, including: mandatory incident reporting; independent evaluators embedded in frontier labs; mandatory early model access for some government agencies such as ASD; a risk-based framework covering model evaluation, safety testing, cybersecurity, transparency, accountability, human oversight and ongoing monitoring. | | The letter also has recommendations for building Australian "sovereign" capability. The full report will be presented to the Parliament before the end of November. | | Trendline: AI scientific breakthroughs carry on. In our last brief, we talked about the solution to the Navier–Stokes Millennium Prize problem. One month later, OpenAI reported solving hundreds of additional open mathematical problems, leading some to call it the most significant day in mathematical history. In the meantime, Anthropic has increased efforts in biology by introducing a new life sciences research group, which then claimed the discovery of a novel enzyme system. |
| |
|
| | Question of the month | | The JSC proposes independent evaluators embedded inside frontier labs with "employee-like access." Supporters say this gives earlier warning than any after-the-fact reporting rule can. Sceptics point to "capture" of the evaluators reducing their independence, increasing security risks, exposing IP, and that it would not be useful given the most risky activities like development and training are likely to be done in other countries. Would such an approach work to increase the safety of frontier AI models? And could Australia enforce any of it? | | Reply to share your answer → | |
|
| | Worth your time | | Companies that get hacked face stricter laws than firms unleashing rogue AI: An article in The Conversation on the liability gap, arguing for an AI "duty of care" beyond online safety, registration of large AI providers and mandatory reporting rules for AI breaches. | | Problems in 22 scientific fields had solutions hiding in plain sight. An AI has found them: A physicist led development of an AI tool to borrow techniques from one research field and apply them to unsolved problems in others, producing 36 papers across 22 disciplines, from linguistics to health policy, in a few weeks. | | Early rogue AI agent activity and attempts to hack found on urlquery.net: Transluce, a non-profit research lab, provides an independent forensic account, with downloadable data, and details on how agent hacking attempts were discovered. | | Speech at the Sydney Trust and Safety Festival: Assistant Minister Andrew Charlton on frontier AI, including why voluntary codes "fail the incentive test." | | How we will do better for Australia: OpenAI’s own account and apology, worth reading as it has more details than previously available (though of course from OpenAI’s perspective only). | | Math 2.0: Terence Tao, a globally leading mathematician raised in Australia, on what his field should keep as AI floods it with results, and a call for a change of pace. |
| |
|
Media enquiries For all enquiries (including media, speaker, education, advisory and research), please email info@gradientinstitute.org and address it to Sarah. |
|
|