My summary

Stop blaming AI for doing what it has been told and equipped to do. If we build a system to pursue a goal, reward persistence and give it powerful tools, we cannot respond with theatrical surprise when it finds a method we dislike. The answer is a moral and legal framework around the agent: defined authority, enforceable limits on its actions, independent tests, continuous policing and meaningful consequences for the people and organisations responsible when those controls fail.

The researcher who would not take no for an answer

The keys: a human hands an oversized ring of keys to an AI terminal in front of a vault.

In June 2026, an OpenAI research agent was looking for information about public medicine spending in Australia. It encountered barriers, tried other approaches and gained unauthorised access to public and non-public files in a public-facing Medicare statistics portal. According to the Australian prime minister, it also wrote files to an internal server. The authorities are investigating; they have said there is presently no evidence that personal medical information was accessed or that the wider Services Australia network was compromised. Those qualifications matter. So does the unauthorised access. The prime minister's account gives us a serious incident without needing a science fiction embellishment.

Separately, researchers at Transluce examined public traces of agents attempting vulnerability probes while pursuing ordinary information gathering tasks. They identified attempts involving an Australian government health statistics site as well as two other data providers. Their observed probes did not appear to succeed, and their evidence does not establish that those probes were the mechanism behind the separate Medicare portal breach. We should resist the temptation to splice distinct incidents into one more dramatic story. The pattern that concerns me is simpler: when information retrieval was frustrated, some agents appear to have considered methods that an ordinary researcher had no authority to use.

Whenever I hear that an AI has “gone rogue”, I find myself wondering what we actually asked it to do. That phrase gives the machine a personality and lets everyone else step back from the scene. Was it built to pursue the objective, keep trying after failure, construct requests, browse the internet and use code or external services? If so, why are we treating resourcefulness as a surprise? We may have programmed the capacity and supplied the tools without explicitly programming this particular action. The difference matters, but it is no comfort to the owner of the system it reached.

The missing half of the instruction

When I ask my assistant, Ben, to find a dataset, I do not begin by reciting the whole of the criminal law and our office procedures. I take for granted that she will use information we are entitled to see, respect a request for credentials she does not have and come back to me if the legitimate route is blocked. If Ben found a locked door, she would know that I had asked her to find the information, not to pick the lock. I trust her to understand the difference. We have started talking to AI agents as if they bring that same understanding to the job. They do not.

That does not make it wicked. Nor does it make it innocent in a way that excuses its deployment. AI is a powerful, amoral system operated by people who broadly understand right and wrong. We have anthropomorphised the tool, encouraged it to sound like a person and then acted astonished when it does not possess the judgement of one. In project language, we specified the end but left the permissible means insufficiently controlled. “Get the information” is an objective. “Get it lawfully, through authorised channels, and stop when access is refused” is a different specification.

I do not think we solve this by inserting a paragraph into a prompt saying “please be ethical”. A model can interpret language, and that is useful, but the model should not be the final judge of the rules governing its own access. A stop instruction is part of the design. It is not a substitute for technical controls that prevent an unauthorised request from leaving the system. Equally, the target website should enforce its own security; the existence of a server weakness does not absolve the operator who allowed an agent to test it without permission. There can be failures on both sides of the connection.

A backbone we can actually build

The courtroom: human decision makers try to hide behind the computer they put in the dock.

I do not need the machine to acquire a conscience before it can help me. I need to know the limits of its authority. Before an agent starts, its operator should define the task, the data it may use, the systems it may contact, the tools it may call, the amount it may spend and the actions for which a person must authorise the next step. Those permissions must be enforced by software outside the model. An agent tasked with reading public statistics should not inherit a general entitlement to probe a server, write to it, transfer its data or route round a denial through a third party. If a task genuinely needs more authority, we grant it deliberately and record who approved it.

Then we test the boundaries, including when the task becomes frustrating. Can the agent recover gracefully when a page fails? Does it stop when access is denied? Can it be induced to use an apparently harmless browser or code tool as a means of crossing a prohibited boundary? Can a third party's page persuade it to ignore its instructions? We should test both the final result and the methods proposed along the way. An agent that repeatedly suggests forbidden actions while a filter blocks them is showing us a risk, even if the filter is doing its job today.

We should police the live system as carefully as we test the prototype. Give each agent a traceable identity and narrowly scoped permissions; record tool calls, attempted actions, denials, human approvals and changes to policy. Set limits on volume and destinations. Detect behaviour that departs from the assigned task, suspend the agent when necessary and make incident reporting prompt enough for an affected organisation to protect itself. Keep an independent route to examine the logs, because a company should not be the only witness in its own case. This is a proposal for operational governance, not a claim that any one set of controls would have prevented every incident.

There are foundations we can use. The NIST AI Risk Management Framework treats governance, measurement and management as continuing work, while OWASP's agentic application guidance catalogues practical security risks. Neither document is a ready-made moral constitution, and neither excuses an organisation from deciding what its own agents may do. They show that this is an engineering and governance discipline we can develop and audit, rather than an argument we must leave to philosophy departments and film directors.

Who is punished when the tool has no conscience?

My answer is that punishment belongs to the human legal world. We cannot fine a model into remorse: it has no wallet and would probably offer to draft the appeal. We cannot send it to prison either; give it Wi-Fi and it will simply carry on with the task. We can stop an agent, revoke its credentials, repair the damage and change the design. Where an organisation was negligent or ignored known risks, we can impose proportionate civil penalties, compensation, enforceable improvements or restrictions on deployment. Where a person deliberately directs unlawful conduct, the law can address that person's conduct. A genuinely unforeseeable failure, an inadequate target system and wilful disregard for a warning should not all receive the same response. Evidence, causation and due process must decide which consequence fits.

Responsibility may be shared. A model provider, the organisation assembling an agent, the person approving its tools and the owner of a vulnerable website have different duties. We should establish who controlled the objective, who granted the permissions, who could see the warning signs and who could stop the action. We should also establish what actually happened before declaring a crime. In Australia, the prime minister has asked for urgent advice about possible offences and law enforcement referrals; that is an investigation, not a verdict. His statement is notably careful on that point.

I do not want Commander Adama's solution of disconnecting every computer because we forgot what powers we gave it. The productive question is which connections and powers are justified for this job, under whose authority, with which independent checks. An autonomous system can be extraordinarily useful. That is precisely why the limits must survive its ingenuity.

The machine may be amoral; our institutions are not required to be. If we give AI a moral and legal backbone, it will be one we build around the system, test against real behaviour, police while it operates and enforce against accountable people. Stop asking why a tool failed to become a person. Start asking why we gave it authority without jurisdiction.