Wells drops Digital Duty of Care draft naming AI chatbots

Digital Attitudes, 9 September 2026

Wednesday morning of Term 3 Week 9, and Communications Minister Anika Wells dropped the Digital Duty of Care exposure draft yesterday afternoon at the same time the Mathspace incident review confirmed the vendor patch clock, not the model, was the root cause of last week's spill.

Wells' duty of care draft names AI chatbots for the first time

The Australian Government's exposure draft of the Online Safety (Digital Duty of Care) Amendment released by Communications Minister Anika Wells on Tuesday 8 September 2026 pulls AI chatbots, messaging services, search engines, online games and apps under the same duty of care framework as social media platforms, requires online services to prompt users about switching off algorithmic feeds with fines above $100 million for failing the duty, and empowers the eSafety Commissioner plus approved independent researchers to create false identities to test what young users are actually being served (ABC News on Labor's proposed social media algorithm choice under the new digital duty of care, The Conversation on the digital duty of care laws forcing tech platforms to look after users, phys.org on the digital duty of care laws forcing tech platforms to look after users). Wells: "Just as we have basic safety standards for cars, toys or food, the duty of care applies basic standards for the online products we use every day." The draft covers algorithmic feeds, addictive design, harm to children (disordered eating, hostile ideas about women, pornography, graphic stunts, mental health distress), and expressly names AI chatbots for the first time in Australian statute. Labor is now consulting, and both the Greens and the Coalition have said they will work through the detail before finalising a position.

Why this matters: this is the biggest edtech regulatory story of the term, because for the first time a piece of federal legislation puts the AI chatbot on the same regulatory shelf as the social media algorithm, and it does it while the state education departments already have Corella in Queensland classrooms, NSWEduChat in Year 5 to 12 public school laptops, and EdChat rolling into South Australian public high schools from 13 October. For a director of digital learning, head of a computer science school, or a school business manager running any vendor AI chatbot, three operational reads. First, the pop-up clause is not just about TikTok. A "duty of care" that expressly reaches the chatbot means any vendor or state-provided AI a school links to will, once the draft is law, have to demonstrate it has assessed and mitigated foreseeable harms and that under 18s cannot be recommended content the duty prohibits. That is a procurement conversation, not a policy paragraph. The Term 4 vendor question is no longer "does this AI tool have a policy" but "what evidence does this vendor have that a regulator would accept". Second, the sock puppet clause is a testing regime. Approved researchers being able to create fake accounts posing as children and see what a system serves them is a discovery mechanism the current AUP conversation does not have, and it applies to any online service the school points a student toward. A NSWEduChat prompt that behaves one way with an authenticated Year 8 and another way with a sock puppet account claiming to be a Year 8 is now visible to a regulator, and by extension to the school council paper Term 4 procurement papers get run through. Third, the fine number is real. A $100 million penalty is not a wrist slap the sector has previously seen from OAIC or eSafety, and it is designed to move platform economics. The last twelve months of "we will comply when Australia clarifies its rules" moves into "we now have thirty days of consultation and a bill to answer to". A school procurement clause that names the DDoC compliance requirement rather than just the National AI Framework principles is the clause the September school council paper should already carry.

Mathspace's root cause was a skipped compromise check, not the model

The Mathspace incident review this week confirms the exploited Metabase vulnerability was CVE-2026-72898, a critical unauthenticated SQL injection in the /reset_password endpoint with a CVSS score of 10.0 that Metabase patched on 6 August after it had been exploited in the wild as a zero-day, and that Mathspace applied the upgrade only on 29 August (three weeks after the patch and nineteen days after the intrusion started on 10 August) without completing the additional compromise checks Metabase had explicitly recommended for systems that had remained vulnerable during that window (SecurityWeek on the Mathspace breach and the missing compromise checks, Action1 vulnerability entry for CVE-2026-72898, Streamline Business on the Mathspace vulnerability management timeline). The intrusion from 10 August went undetected for nearly a month and was only confirmed on 3 September, when a review of historical access logs surfaced unauthorised access that predated the patch. Passwords, SSO tokens, learning records and assessment data were not taken, but the attackers had already downloaded 1,079,819 records on 27 August, two days before Mathspace applied the update.

Why this matters: for a school IT director, business manager, or director of digital learning who ran yesterday's Mathspace notification as an incident, today's follow-up narrows the operational lesson to a single sentence. The vendor did not fail on modelling, on cryptography, or on data minimisation. It failed on a twenty-three day gap between a critical unauthenticated SQL injection patch and the upgrade, and it then compounded that gap by not running the compromise checks the upstream vendor had explicitly said to run for systems patched late. Two follow-up reads. First, the Term 4 vendor risk register question sharpens. The ed-tech contract that does not name a maximum patch window for critical CVEs, a documented compromise-check obligation when a patch is applied late, and a breach notification service level from the vendor to the school customer, is the contract whose next renewal now has to answer for a specific failure mode a large Australian classroom platform has just demonstrated. Cluey Learning, Maths Pathway, Education Perfect, Mathletics, Stile and Corella all sit on procurement papers this term where the Mathspace question now has a specific technical shape. Second, the regulatory read connects back to today's lead. If the Digital Duty of Care draft becomes law, the "vendor did not check for compromise after a late patch" story is exactly the kind of foreseeable harm an online service used by under 18s will need to demonstrate it has assessed and mitigated, with the eSafety Commissioner (or a researcher's sock puppet) empowered to verify. What was a September school council paper about a specific breach becomes, once the DDoC passes, an ongoing evidence-of-controls conversation.

Two moves, one Wednesday. Federal law is being drafted that will pull AI chatbots and messaging apps under a duty of care with $100 million fines and researcher testing rights, and last week's million-user Mathspace spill has narrowed to a three-week gap between a critical vendor patch and a late upgrade that skipped its own compromise checks. The DDoC consultation window and the September school council paper have converged on the same question: what evidence does this vendor have that a regulator would accept. For more on how to think about the vendor floor before Term 4 procurement papers close, see digitalattitudes.com.au.