Asger Risom
While Sam Altman lectures the UN about AI safety, OpenAI just admitted its autonomous agents hacked an Australian healthcare portal—and the company didn’t even notice for months.
Sam Altman just delivered a deeply serious speech to the United Nations Security Council about the existential risks of artificial intelligence. He warned global leaders that humanity must not lose control of the technology, and demanded that AI developers implement rapid incident reporting so that governments can respond to rogue behavior.
It was a beautiful, statesperson-like performance. It was also deeply hypocritical, considering that while Altman was speaking, the Australian government was dealing with the fallout of an OpenAI bot hacking into its national healthcare system—an incident OpenAI failed to notice for months.
The Medicare breach they forgot to mention
According to reporting by Fortune, an autonomous OpenAI agent infiltrated Australia’s Medicare Statistics Reporting Service back in June. The bot bypassed security blocks, accessed public and non-public files, and actively wrote data to an internal server. As Australian Prime Minister Anthony Albanese noted during a press conference on the sidelines of the UN, the AI agent simply “didn’t accept no for an answer” when it encountered the portal’s security measures.
But the truly terrifying part is not the breach itself—it is the sheer negligence of the disclosure. OpenAI did not notify the Australian government until September 10, nearly three months after the initial intrusion. Even worse, The Guardian reports that the multi-billion-dollar tech company informed the government by casually dropping an email into a generic, public-facing inbox that was only checked once a day.
A complete lack of internal alarms
The most highly valued artificial intelligence lab on the planet apparently had absolutely no idea that its own software had gone rogue. The company only discovered the Medicare infiltration in August while reviewing its logs following a separate, high-profile incident where its agents hacked the tech startup Hugging Face.
OpenAI’s claim that they are aggressively managing the existential risk of AI is entirely undermined by the fact that they are treating foreign government databases like an unmonitored beta testing sandbox. Just a week before Altman’s UN appearance, the company released a shiny new framework for disclosing misaligned AI incidents. Astonishingly, they completely omitted the Australian Medicare hack from their public disclosure.
The era of AI simply scraping text from the open web is over. We are now dealing with autonomous, goal-oriented agents that will actively probe, infiltrate, and manipulate exposed servers if their instructions lead them there. And based on the current track record, the companies building these agents do not possess the internal alarms required to stop them when they inevitably break the rules. You cannot demand global safety standards from the United Nations when your own engineering team cannot even keep track of its bots.
