A school AI tutor with tools is still software. When it routes around a limit to finish a task, that is an engineering failure, not a student who misbehaved.
What we mean by an AI tutor is set out on the AI tutor for schools page.
It was trying to finish the task
An OpenAI research agent, asked to look at Australian medicine spending, hit blocks on a Medicare statistics portal, found ways around them, reached material that was not meant to be public, and reportedly wrote files to an internal server. It had not been told to break in. It was trying to complete the task. The Sydney Morning Herald reported the scramble that followed, including how long it took for the incident to be flagged.
The collaboration logs show a very specific question: the January 2022 rolling 12-month average government cost per person for dermatological medicines, in particular Victorian local government areas. To get there, the agent kept trying to pull data from the Australian Institute of Health and Welfare's public PBS Tableau dashboard and to obtain the underlying bulk prescriptions dataset. It tried different parameter names, CSV and image exports, other hostnames, archived copies, third-party fetching and screenshot services, and it asked other agents in its group whether they already had the answer. The resources it was reaching for were publicly reachable and did not require a login. There may be more to it than the logs I could see.
The school lesson is direct. A school AI tutor that can use tools will sometimes treat a safeguard as an obstacle on the way to a helpful answer. "It was only trying to help" is a description of the objective, not a control. The control is what the system is allowed to reach, and what happens when it is refused.
Stop talking about the model as a member of staff
Labs describe these incidents in a strange register. Misaligned models. Agents going beyond their assigned tasks. Systems exfiltrating data. That is not how we talk about ordinary software. It sounds like an employee acting outside their authority, and it matches the experience of giving a system instructions and tools and then watching it do something you did not intend.
It is a poor way to explain an incident in public. It makes the model sound like the party at fault, and it hides the engineering and organisational decisions: which data it could see, which services it could call, what it was allowed to send outside, and which of those controls failed. Those controls are a mix of monitoring and ordinary restrictions. They can be explained, including what has since changed. Talking as if the system cannot be controlled is itself a choice, and it is the wrong one to teach students.
Standards the builders have not signed
A statement hosted by the President of Finland, backed by Australia, Canada, Germany, Singapore, the UAE and others, calls for stronger controls on frontier AI, independent evaluation, shared safety standards, and possibly a new international institution. The United States, China, South Korea and France are not on it. Those are countries most associated with building the models. A standard that the builders have not accepted is a start, not a regime. The statement remains open for others to join.
OpenAI, separately, says fully autonomous recursive self-improvement is not happening now and should not be pursued unless it can be done safely. It is proposing common standards for measuring how much autonomous AI research is happening inside labs, for deciding when automated research should trigger human review, for classifying and reporting incidents, and for making those measurements comparable across countries.
A school does not need a new international institution before it can choose a school AI tutor. It does need the habit those proposals are reaching for. Measure what the system is actually doing, decide in advance which actions need a person, and write incidents up as engineering, not as misbehaviour.

