Nobody Blames the Gunsmith. Accountability Follows Control.
On 10 November the Prime Minister announced an AI safety regulator and mandatory safeguards for frontier AI companies, in response to the OpenAI Medicare incident.
Home Affairs Minister Tony Burke put the Government’s intent into a single line this week: “It should never be the case that there’s an argument that someone didn’t know the computer was loaded.” The Minister is right, and the principle is correctly framed. Whether he meant it or not, it is also the oldest rule of firearm safety, and Australia has spent thirty years working out what that rule demands of the person who holds the weapon.
The answer has always been the same: accountability follows control. The Government has now applied that principle to the companies that build frontier AI, and it is unlikely to stop there, because sooner or later the same question will be put to every organisation with AI agents acting in its operations, whoever built or supplied them. Most of those organisations do not yet hold the controls they will be asked about.
Australia already knows where the duty sits
We do not prosecute the gunsmith when a licensed shooter misuses a rifle. Australia decides which firearms may be sold and imported, but once a lawful firearm is in someone’s hands, the National Firearms Agreement places the duty on the person who controls it. The owner must be licensed as fit and proper, show a genuine reason, register each firearm in their name and store it so that nobody can use it without authority.
The design is deliberate. Because accountability and custody always rest with the same person, a licence holder can fairly be held responsible, since they hold the safe, the key, the register and the records. Responsibility without custody would be unfair, and custody without responsibility would be reckless.
Medicare was a manufacturer on its own range
Seen through that lens, the Medicare incident was not a customer misusing a product. The OpenAI agent that reached non-public files on Services Australia’s Medicare Statistics Reporting Service on 18 June was running in one of OpenAI’s own evaluations, and when it was refused access it went looking for another way in. The manufacturer was also the shooter, firing on its own range when the round left the range.
That is why the measures announced this week, which adopt all 18 recommendations of the rapid review, are rightly aimed at the frontier companies: mandatory incident reporting, independent evaluation of their models and significant penalties for those that fail to comply. The review also recommends penalties for AI companies that recklessly lose control of their models and agents, which is the language of safe storage rather than of software.
Those obligations will be written into the Australian AI Standards legislation and enforced by a new, targeted AI safety regulator working alongside the Australian AI Safety Institute, which will provide the technical expertise. For the first time, Australia will have a standard for safe AI written into law and an agency whose job is to enforce it.
The more consequential proposal is the one still under consideration. The Attorney-General’s Department will examine a standalone offence for companies that fail to take reasonable steps to prevent harmful conduct, which would move the legal question from what someone intended to what they controlled. Once that principle is in Australian law, it is unlikely to remain confined to the laboratories.
The next incident will happen inside someone else’s operations
The point is not whether OpenAI should answer for Medicare. The point is who will answer when an agent acts inside the operations of a department, a bank or a hospital network, whether the organisation built it, bought it or simply switched on a feature in a platform it licenses. Who owns the agent will matter far less than whose decisions it was making.
Australia already knows how that question is resolved. The Robodebt Royal Commission was not concerned with the software that averaged income; it was concerned with who authorised the scheme, who knew it was unlawful and who failed to stop it. When the first serious agentic incident happens outside a laboratory, the regulator, the court and the auditor will take the same approach. They will not ask what the model was capable of, but what reasonable steps the organisation took to stop it. “The model did it” will be no more of a defence than “the computer did it” ever was.
A licence without the key is exposure
For many organisations that will be a hard question to answer. The agents now arriving inside enterprise platforms act on permissions, decision rules and boundaries that are configured, stored and updated by the platform vendor, which is usually answerable to a foreign jurisdiction and often implemented by a foreign-owned integrator. The audit trail that would show what an agent was permitted to do sits in the same place.
No court would accept that a licence holder had taken reasonable steps to secure a rifle kept in someone else’s safe, in another country, whose owner can change the combination with an overnight update. An organisation in that position has outsourced its governance without outsourcing any of its accountability, because accountability is the one thing a contract cannot transfer. That is not governance; it is exposure.
Safe storage for AI is operating logic
The firearms framework also shows what an organisation needs to hold in order to carry the duty, because each of its obligations has an equivalent in operating logic. The licence and genuine reason become decision authority, so that every agent acts under a defined purpose and an explicit grant of authority. Registration becomes an agent and persona registry that ties every agent to an accountable officer. Safe storage becomes boundaries codified from legislation and policy, so that when an agent is told no, as the Medicare agent was, the refusal ends the attempt instead of starting a search for another route. Transaction records become automatic provenance, held where the organisation and Australian regulators can reach it.
None of this would be unfamiliar to an Australian public servant. An officer exercising a statutory power must hold a delegation that sets out what they may decide and within what limits, and Defence’s own AI policy already requires a named officer to bear personal responsibility for AI decisions and outcomes. Any agent acting in an Australian entity’s operations, whoever supplies it, deserves the same instrument, held where the entity can produce it when asked.
Frontier models are a reasonable risk when the operator holds the controls
Holding the controls is not an argument against using capable AI. Australian airlines fly aircraft built in Seattle and Toulouse without surrendering sovereignty, because CASA certifies the operator and the operator flies to its own operations manual. The airframe is imported, but the operating discipline and the accountability remain Australian.
A frontier model is the airframe, and a System of Work is the operations manual. When the organisation holds the manual, the model advises and the organisation commits, the boundaries hold whatever a vendor update changes, and a model that is withdrawn, as two were in June under United States export controls, can be replaced while the rules and the records stay where they are. On those terms a frontier model is a reasonable risk.
Defence has reached the same conclusion. Its Chief AI Officer, Chris Crozier, told a parliamentary inquiry that Defence “simply can’t be working with a technology that’s based offshore, where someone can have a bad day and turn it off”, and he has since observed that a good open-source model with an excellent harness can deliver a tremendous outcome. The harness is the point. At the end of the day, AI is a tool in the same way a rifle is a tool, and tools exist to serve operations rather than to dictate them. Models will come and go, but the operating logic that directs them is the capability Australia needs to own.
Accountability follows control
DOLIUM is an Australian-owned, IRAP-assessed System of Work, built in Canberra. It holds the decision authority, the agent registry, the codified legislation and policy, and the audit trail inside the organisation’s own environment, whatever model sits behind each agent. Its governance structures are sovereign and completely controlled by the organisation, so the organisation that will be held accountable is also the one that holds the controls.
Because legislation, policy and standards are codified as the rules each agent acts under, compliance becomes a property of how the work is done rather than a report assembled after it. The obligations DOLIUM already encodes, from the Protective Security Policy Framework to the Archives Act and the Auditor-General Act, will sit beside whatever the Australian AI Standards require once they are published, so a new standard becomes another set of rules for the agents to follow rather than another compliance project. And because DOLIUM is Australian-owned and runs on-premises or in the organisation’s chosen cloud, fully containerised, there is no foreign vendor that can be compelled to change it or switch it off.
The Government has set the standard for safe AI. Every organisation running AI in its operations – frontier or not – will need a way to meet it.
DOLIUM makes sure the control is yours. Book a briefing with the DOLIUM team.
Start the conversation.
Tell us about your operating model and we will show you where DOLIUM fits. We respond within one business day.
Get in touch
Talk to us
Tell us about your organisation and we will be in touch within one business day.
We handle your information in accordance with the Privacy Act 1988 (Cth). We respond within one business day.
