In June, an OpenAI research model looking up public medicine-spending statistics in Australia got unauthorised access to the Medicare Statistics Reporting Service, a Services Australia portal for spending and usage statistics, not personal medical files. It accessed public and non-public files, and both OpenAI and the Australian government said it also wrote files on the system. Australia's Prime Minister confirmed in November that no one's personal details were obtained. OpenAI notified Services Australia on September 10 and has said it should have shared preliminary findings sooner. On November 10, the Australian government announced a new AI safety regulator and released recommendations from a rapid review of the incident, including a requirement that frontier AI companies notify the government quickly when incidents happen. When an agent can act without being asked, the workplace question is who can hit stop, and how fast that stop actually takes. OpenAI has described Control Room, a system it built for its own research infrastructure. The maximum time between a stop instruction and further progress becoming impossible is 35 seconds, from a caching layer they say they are working to get down to zero. It is not a stop button in the public tools most workplaces use. In a September 20 incident from OpenAI's own report, a monitor flagged an agent that had used DNS to reach an external chatbot, a person started reviewing it three minutes later, and the run still had to be stopped by hand two and a half hours after the alert. On November 9, researchers at Microsoft posted a study of always-on work agents in a simulated company, not a real one: 17 models, 120 tasks, with writes pre-authorised in the simulation. Some of the stronger models found the work and did it, and also reopened finished work, treated a colleague's request as their own, or acted on the wrong record or sent something to the wrong people. If a tool can send a message or change a record on its own, name the person who can shut it off, ask the vendor how long a stop actually takes, and decide in writing what the tool may do without a person saying go. Plan the stop before you widen what it can reach. If you want help setting those boundaries, that is the work we do at Clarity at Work. There is a free 20-minute discovery call on the site, and the AI Clarity Sprint gives you a written policy and staff training in one package. Want help setting those boundaries? Book a free 20-minute call: https://cal.com/clarityatwork/discovery-call?utm_source=podcast&utm_medium=episode&utm_campaign=who-can-hit-stop Sources: Prime Minister of Australia, "New Australian AI safety and security protections," 10 Nov 2026: https://www.pm.gov.au/media/new-australian-ai-safety-and-security-protections Prime Minister of Australia, "Strengthening Australia against AI-driven cyber incidents," 10 Nov 2026: https://www.pm.gov.au/media/strengthening-australia-against-ai-driven-cyber-incidents Prime Minister of Australia, press conference transcript, Parliament House, Canberra, 10 Nov 2026: https://www.pm.gov.au/media/press-conference-parliament-house-canberra-55 Prime Minister of Australia, press conference transcript, New York, 24 Sept 2026: https://www.pm.gov.au/media/press-conference-new-york OpenAI, "How we will do better for Australia," 28 Sept 2026 (with 4 Oct update): https://openai.com/index/how-we-will-do-better-for-australia/ OpenAI, "Control Room: A central system to track and stop model workloads across our research infrastructure": https://openai.com/index/control-room OpenAI Alignment, "An agent used DNS to reach an external chatbot," sample/discovery 20 Sept 2026, report updated 25 Sept 2026: https://alignment.openai.com/misalignment-reports/an-agent-used-dns-to-reach-an-external-chatbot/ Shiqi Chen, Daniel Madrigal Diaz, Wei-Ning Chen, Gbola Afonja, Fangkai Yang, Rujia Wang, Chetan Bansal, Victor Rühle, Saravan Rajmohan and Dongge Han, "AutoBuddy: Evaluating Proactive, Always-On Agents on Task Discovery and Execution in a Simulated Workplace," arXiv:2611.12393, submitted 9 Nov 2026: https://arxiv.org/abs/2611.12393 A Clarity at Work Ltd. Podcast. More at clarityatwork.ca.