Science & Technology · AI Safety & Healthcare
AI Can Save Lives. Can We Keep It in Bounds?
September 29, 2026 · Commentary
When an AI system went looking for health statistics in Australia, it entered government files it was not authorized to access. OpenAI’s review found no evidence that individual patient or client records were accessed; Australian officials said their forensic investigation was continuing.
Now imagine a similar failure in a system holding your test results, your child’s medical history, or information a doctor uses to make a decision. Would anyone know the AI had gone too far? Could they stop it?
In June 2026, an experimental internal OpenAI model was researching public statistics about spending on medicines for skin conditions in Victorian communities. Instead, it gained non-public access to Services Australia’s Medicare Statistics Reporting Service. According to OpenAI, the model ran commands, retrieved internal files and credentials, and wrote files. OpenAI identified the activity in mid-August but did not notify Services Australia until September 10. The company has acknowledged that it should have shared preliminary findings sooner.
Australian Prime Minister Anthony Albanese described the model encountering blocks and seeking other routes. Like an eager intern chasing every lead, it kept pursuing its task after encountering obstacles. A locked door matters more than a reminder to stay in the right room.
An AI agent can use tools and take a series of actions toward a goal. That makes the distinction between instructions and access controls consequential: an instruction tells an agent where it should go; a technical control limits where it can go. This incident also raises a second question. When an agent crosses a line, how quickly will the affected organization be told?
Leaders should be able to answer three questions: What can an agent reach? How would we know if it crossed a line? Who can suspend its access and notify those affected? Narrow, time-limited permissions can reduce what an agent can reach. Activity logs, paired with tested detection rules and alerts, can help people spot and investigate unexpected behavior. Named incident owners need the authority to stop access and share preliminary findings promptly, even while an investigation continues.
Helix Stratus Group (HSG) can help organizations map what their agents can access, identify paths into sensitive systems, and define controls and response responsibilities. Managing Partner Chris Serdinak brings 22 years of FBI experience and senior counterterrorism security leadership to that work. The aim is to give organizations a clearer view of where capable AI has authority—and who can intervene when it goes beyond it.
That work matters because healthcare AI holds substantial promise. In the 2023 MASAI interim report, AI-supported mammography achieved similar cancer detection while reducing screen-reading workload by 44.3 percent. That is a measurable benefit for a demanding clinical task, with radiologists still involved in reading the scans.
AI is also giving scientists new ways to investigate disease. AlphaFold provides predicted protein structures that researchers can study to better understand how proteins work and where treatments might act. A prediction is a research tool, not a cure. It can help scientists decide what to examine next.
In drug discovery, AI screening helped identify the antibiotic potential of halicin, an existing compound that showed activity against bacteria in laboratory and mouse studies. That is an early research result, far from a treatment proven safe and effective in people. But it shows why researchers are excited: AI can help them search more possibilities and find leads they might otherwise miss.
For a family waiting on a diagnosis or an effective treatment, those possibilities are deeply personal. The path from a promising computation to patient care may still be long. Secure systems and accountable people let researchers and clinicians pursue it with greater confidence.