| | Making sense of… | | AI agents interacting across organisational boundaries | | The Australian AI Safety Institute has published its first report – one that was written by Gradient Institute. The report develops a technical framework for the risks, controls and governance of AI agents performing tasks across organisational boundaries. Companies and individuals are giving autonomous agents longer and more complex tasks. As a result, those agents increasingly interact with one another, which creates new risks. And because the interactions span organisations, the resulting safety issues extend beyond what any single organisation can control on its own. The report’s analysis – covering singular, federated, and open environments – is intended to help policymakers, researchers, and practitioners assess and manage the risks of these complex, varied, and often unpredictable AI deployments. | | Read the full analysis → | |
|
| | On the radar | | 1. Agents are doing unsanctioned things, with strategies that involve impersonating people and leaving notes for each other. The UK AI Safety Institute ran a routine cyber evaluation and reported that, in one strategy to win the cyber challenge at hand, an agent impersonated a couple of human accounts on GitHub, posting messages to invite the maintainer of a software repository to accept malicious code. | | 2. Around the same time, OpenAI revealed that prior to their models escaping and attacking Hugging Face (see the previous Gradient Brief), models under training built a shared message board to discuss approaches to hacking and cheating. As we go to digital press, OpenAI has released its technical report into the incident, and METR and Redwood Research have released their independent investigation. | | 3. Australia’s institutions are organising different initiatives for managing the benefits and risks of AI. Fergus Hanson – who specialises in diplomacy, cyber matters, and national security – will head the new Office of AI in the Department of PM and Cabinet, NSW has also elevated AI by creating a new Office of AI in the Cabinet Office leading to the need for a name change of their existing one, South Australia has announced an AI Royal Commission, the Federal Parliament has announced a Joint Select Committee on AI and NSW has published its Operational Policy. | | Trendline. A July report from DEWR showed that from late 2022 to Feb 2026 there was no large disruption to employment in Australia due to AI. The report found the labour market for young workers and young graduates had not shown broad deterioration. The Stanford Canaries dashboard indicates the opposite for young workers in the United States. Two possible explanations are: that Australia’s adoption of AI is lagging and we are therefore witnessing the US situation in 2024, or that the two labour markets have such different characteristics that the American result does not apply. DEWR has established their multi-dimensional framework for ongoing monitoring and this will be an important set of signals to track. |
| |
|
| | Question of the month | | Can oversight fatigue be solved with better training for overseers and better design of the AI system, or is it just an unavoidable cost of automation? | | In 1983, Lisanne Bainbridge’s Ironies of Automation paper pointed out a paradox that still holds today: the more reliable a system becomes, the harder it is for a human to stay alert while monitoring it. This same dynamic plays out in human oversight of AI systems, including behaviours such as overreliance, automation bias and rubber-stamping. It raises the question of whether we can train or design our way out of these behaviours, or whether it’s simply a risk we must accept when humans oversee increasingly capable AI systems. We’re curious what you think: we read every answer. | | Reply to share your answer → | |
|
Media inquiries For all enquiries (including media, speaker, education, advisory and research), please email info@gradientinstitute.org and address it to Sarah. |
|
|
