Csiro Warns Farmers Over Ai Risks As Systems Take Control

Australian farmers are increasingly turning to artificial intelligence (AI) to boost efficiency and manage complex operations, but experts warn that handing decision-making authority to autonomous agents could trigger costly mistakes.

or signup to

The dangers recently attracted fresh public attention following revelations that an OpenAI agent gained unauthorised access to a Services Australia Medicare statistics reporting portal earlier this year, accessing non-public information, while researching public medicine spending.

The incident raised questions across the economy the risks when AI agents interact with systems that were not designed to prevent every action an agent might attempt.

For agriculture, the issue is relevant to an industry increasingly reliant on connected machinery, digital records and software to inform or make decisions.

AgForce general president Shane McCarthy said the attraction of AI tools was clear for producers under pressure to make production gains.

"We run on such bare margins as it is now, and we don't control the price that we get," he said.

"We've got to be able to increase efficiency, and smart technology and AI are some of the few ways we can do it."

Macquarie Valley cotton, beef and sheep producer Andrew Gill recently told The Land that when he uploaded data from his John Deere tractor to an AI-based farm assistant he is trialling, the system flagged fuel consumption of 17 litres a hectare in a deep-ripped paddock, compared with four litres across much of its work that season.

Mr Gill had not specifically asked the AI to investigate the difference , and the example shows how AI can find useful insights in information a farm already collects.

The risks change, however, when a system moves beyond analysing information and is given authority to act on it.

CSIRO senior research scientist Dr Chandra Thapa said that relationship could change, however, when users delegated authority to an AI agent - or allowed it to not just interpret information or provide advice, but act on it.

"Farms generate large volumes of data from diverse sources, including soil sensors, machinery, livestock records, satellite imagery, financial systems and supply-chain platforms," he said.

"Unlike an AI system that only provides advice, an AI agent could use authorised access to move between services, combine information and complete a sequence of tasks."

That ability could be useful to a farmer who wants a system to bring together crop and weather data, check irrigation requirements or manage orders across several services.

But the ability to act introduces a different risk from simply getting an answer wrong, and giving an agent a goal does not necessarily mean the user has anticipated every action it might take to achieve it.

It was recently reported by the ABC that an Australian man used an AI agent to book a gym class but found it had exploited a vulnerability in the booking system and removed another person from the class.

The agent had not been explicitly asked to remove another person, and the booking software lacked adequate authorisation checks.

Dr Thapa said an agent with access to farm data and connected services could potentially alter irrigation schedules, place orders or approve transactions if the underlying system were compromised.

There are also risks when agents encounter misleading instructions or information while carrying out a task.

AI security researchers have raised concerns prompt injection, in which content embedded in a document, email or website attempts to redirect an agent away from its original instructions.

Whether such an attempt succeeds depends on the system's design, the access it has and the safeguards in place.

The issue comes as data centres are increasingly looking to regional areas , raising concerns competition for electricity, water and agricultural land as demand for AI computing grows.

AusAgritech executive officer Britta Marsh said the distinction between analysing information and acting on it was becoming increasingly important, and permission for AI to analyse farm data should not automatically mean giving it a green light to control machinery, change settings or that data.

"Preparing a report is different scenario from controlling irrigation or chemical application," she said.

Ms Marsh said higher-consequence actions required stronger testing, monitoring, approval and practical ways for humans to intervene.

The same principle applies to information an agent can access.

An agent helping with a purchase, for example, may need access to payment details or financial accounts and users must be across what actions the agent can take with it and when it must seek approval.

Giving an agent broad access simply because it makes a task easier can create risks that are difficult to anticipate.

The National Farmers' Federation voluntary Australian Farm Data Code provides an existing framework for transparency and control over how farm data is collected, used, managed and shared, and sets expectations for service providers handling farm data.

The lobby group points that data generally cannot be legally owned in Australia, making the terms governing its collection, use and sharing particularly important.

The warnings arrive as more growers experiment with AI-powered tools designed to simplify decision-making and reduce administrative workloads, such as summarising chemical or rotation records, drafting compliance reports or comparing input prices across suppliers.

"Members report growers wanting connected tools and more value from data they already collect," Ms Marsh said.

"Trusted integration can reduce duplication, improve decisions and support innovation, with transparent commercial arrangements for both farmers and technology providers."

Incorrect advice can also have practical consequences even when a system has no authority to operate machinery or make purchases.

A landholder who uploads a photograph to identify a weed, pest or disease, for example, could act on an incorrect answer and apply an unsuitable chemical, potentially creating residue, withholding-period or resistance problems - particularly if the model was not trained in Australia.

Handing over authority

CSIRO's Agriculture and Food division said agricultural AI presented risks beyond malicious attacks, because farm decisions were made amid changing seasonal conditions, biological variability and incomplete information.

It said an AI system could produce an unreliable recommendation without being hacked or deliberately manipulated, and farmers and their advisers needed to be able to question recommendations, see what information an AI system had used, compare them with their own knowledge and override them where appropriate.

In short, use AI to support rather than replace human judgement, particularly for decisions with significant operational, financial or production consequences.

CSIRO also said the technology provider also had a role in managing risk and, given how deeply AI was being embedded in software and other applications, that it was impractical to expect farmers alone to mitigate those risks.

Choosing known and trusted providers was one part of reducing them.

There is evidence of cyber incidents disrupting agricultural operations, with Mackay Sugar reporting a cybersecurity issue that impacted operations in June this year, issuing several updates as it responded.

But CSIRO drew a distinction between those established cyber risks and the more specific prospect of AI systems being deliberately compromised to cause harm on farms. It said there was not yet strong evidence of the latter.

The division said the concern over the five years was the potential for AI to become more deeply connected to farm data, software, machinery and other services. Researchers say this is why farmers must understand where AI recommendations come from, how reliable they are and when human judgement remains essential.

For example, if an AI agent misinterprets a dry spell as a long-term trend, it could recommend planting decisions that increase financial risk if acted upon or set an autonomous sprayer to apply the wrong chemical at the wrong rate.

Ms Marsh said industry standards would become essential as AI systems became more deeply integrated into agricultural operations.

"Responsibility must be clear when something goes wrong," Ms Marsh said.

Another issue is what happens when an agent is given a new task, depending on how a tool works, instructions or retained context may still influence what it does .

The ability to run multiple agents or instances also changes the scale of the risk. An individual might perform one action at a time, while automated systems can potentially make repeated attempts or carry out tasks in parallel. That can be useful for routine work, but it can also multiply mistakes or unintended actions before a user notices.

Mr McCarthy said the pace of technological change was creating a challenge for farmers trying to keep up with both the opportunities and the safeguards.

"The tech is probably moving faster than people adapt to it," he said.

"The tech evolves, and then we've got to find safeguards."

National rural affairs writer covering federal politics, agri-politics and life in the regions. Winner of several Clarion Awards, including best news report, the 2025 Australian Council of Agriculture Journalists Star Prize and two National Rural Press Club awards. Send story tips to [email protected]

National rural affairs writer covering federal politics, agri-politics and life in the regions. Winner of several Clarion Awards, including best news report, the 2025 Australian Council of Agriculture Journalists Star Prize and two National Rural Press Club awards. Send story tips to [email protected]

More from Agribusiness

Can farms get hacked? Cyberattacks put nation's food security in firing line

What Labor's crackdown on discretionary trusts means for family farms

Hard-wired: historic fencing brand is agriculture's new 'most trusted' name

Contentious trust tax bill tabled as farmers, small business face major changes

How big names in egg 'cartel' allegedly restricted supplies and blocked rivals

Record yields, biggest canola harvest on record: Australia's 63m tonne crop