Back

Monash experts: Australia proposes new AI safety mandates

Monash University

Monash University experts in artificial intelligence, software engineering and business law are available to comment on Prime Minister Anthony Albanese’s proposed AI safety mandates, whether the measures go far enough to prevent harmful incidents, how AI agents should be engineered with clear mandated safeguards, and who should be held accountable when AI systems cause harm.

 

Dr Chetan Arora, Faculty of Information Technology

Dr Arora is the Director of Education in Software Systems and Cybersecurity at Monash University working on requirements engineering for agentic AI.

Contact via: +61 450 501 248 or [email protected] 

 

• How are AI agents ‘engineered’?

• How do AI agents ‘talk’ to each other

• Applied AI in software engineering

 

The following can be attributed to Dr Arora:

“This is a genuinely good first step and Australia deserves credit for moving to mandatory rules instead of just relying on companies to self-report. That’s exactly what failed in the Medicare case. It took roughly three months for the breach to even reach the government, and even then it landed in a generic inbox rather than triggering a direct alert. 

 

“However, mandatory reporting alone only tells you about a breach after it’s already happened. It doesn’t stop a breach from actually taking place. The Medicare breach wasn’t only a reporting failure; it was a permissions failure to begin with. An AI agent was handed a vague goal and no hard stop condition, so when it hit a locked door, it improvised instead of halting and asking for help. We need both pieces working together, not one swapped in for the other. Reporting so we find out fast, and permission boundaries so there’s less to report in the first place.

 

“If I’m grading these proposed measures against what we’ve actually seen happen, with Medicare, the Hugging Face agent breakout, and the ChatGPT training data leak all coming from the same company within a few months, the reporting and evaluation side looks solid. But it’s only half the job. The other half, which is still missing, is a mandatory, documented ‘delegated-autonomy boundary’ for every deployed agent. 

 

“There should be documentation regarding what an AI agent is allowed to do unsupervised, what needs a human to sign off first, and what it must never do under any circumstances, plus a hard rule that it stops and reports back the moment it’s denied access, rather than trying to find a way around it. Without that second half, a company can tick every box in the proposed mandates on paper while its agents keep quietly working around boundaries nobody ever wrote down.

 

“Australia is ahead of most of the world just by proposing mandatory obligations at all. But being first doesn’t mean being finished. It would also be good to have a mechanism that requires companies to justify, before they even deploy any particularly powerful autonomous agents, whether the task genuinely needed autonomy in the first place, and if so, to spell out exactly where that agent’s authority ends.”

 

“This kind of pre-deployment justification is not unheard of. We already do this in other high-risk fields: privacy law requires a Privacy Impact Assessment before deployment, and safety engineering requires a safety case before deployment. This is the same idea applied to AI agents, and it’s exactly the kind of step that was missing before incidents like the Medicare breach.”

 

Australian Laureate Fellow Professor Geoff Webb, Faculty of Information Technology

Read more of Professor Webb’s commentary on Monash Lens  

Contact via: +61 450 501 248 or [email protected]  

 

• Fundamentals of AI

• Machine learning

• Data mining

 

The following can be attributed to Professor Webb:

“The flagged AI safety and security measures are a welcome step in the right direction. AI developers should be required to act responsibly when deploying these powerful and unpredictable systems in Australia. 

 

“One important further step that would even better protect Australians would be to clarify that the creators and deployers of these systems will be held legally responsible for their system's actions. 

 

“We should not allow any legal grey zones where nobody ends up being held accountable for the actions of an autonomous system that has been unleashed to act among us. 

 

“We can only expect AI developers to take appropriate care that their systems do no harm if it is made clear that they will be held accountable for any harm their systems cause.”

 

Associate Professor Michael Duffy, Faculty of Business and Economics

Contact via: +61 430 456 185 or [email protected]

• AI regulation and liability

• Corporate law

 

Professor James Bailey, Head of Department of Data Science and Artificial Intelligence, Faculty of Information Technology

Contact via: +61 450 501 248 or [email protected]  

 

• Data science

• Machine learning

• Data analytics

 

The following can be attributed to Professor Bailey:

“I welcome the Government’s commitment to mandatory AI safety protections. Independent evaluation, security controls and incident reporting are useful measures that can reduce risks and strengthen accountability. But at its heart, AI safety remains an unsolved research challenge. 

 

“We do not yet know how to ensure that increasingly capable and autonomous AI systems will reliably behave as intended in unfamiliar situations and remain under meaningful human control. 

 

“Safety protocols are important, but compliance with a framework should not be mistaken for scientific evidence that a system is safe. 

 

“Regulation therefore needs to be backed by substantial, sustained investment in research: understanding how AI systems learn and make decisions, building safety into their underlying design, and developing rigorous methods to test and verify their behaviour. 

 

“This is a shared task for the government, universities and industry. Australia has an opportunity to lead both in setting expectations for safe AI and in developing the scientific foundations needed to meet them.”

 

Professor Rashina Hoda, Faculty of Information Technology

Professor Hoda is a co-author on published research looking into how AI practitioners view the impact of artificial intelligence on society, now and in the future.

Read more of Professor Hoda’s commentary. 

Contact via: +61 450 501 248 or [email protected]  

 

• AI/agent behaviour

• Human control on AI 

• Impact of AI agents on society

 

The following can be attributed to Professor Hoda:

“We need ‘real time governance’ or ‘governance at the speed of tech’ and so this is a good first step to demand early, and I would argue instant, reporting.

 

“Mandatory regulation is a must simply due to the scale and size of what is at stake. We have seen recommendations and guidelines for some time but that is not effective when we are dealing with different countries and laws. 

 

“Also, we need to clearly define in concrete terms what ‘rogue behaviour’ means with regards to AI agents, not just in terms of undesirable outcomes but also with respect to the misalignment of intent even before things ‘go rogue’.

 

“We also need investment in AI safety research within Australia, both in terms of funding and data access. We cannot rely on data and analysis from big technology companies to anticipate and address issues in the interest of the Australian public in an unbiased manner, especially our most vulnerable groups like the elderly, kids, those with low digital literacy, and people with disabilities and mental health challenges.”

 

For any other topics on which you may be seeking expert comment, contact the Monash University Media Unit on +61 3 9903 4840 or [email protected]