A day after Anthony Albanese revealed that an OpenAI agent had got into a Medicare statistics portal, OpenAI has admitted that its AI agents misbehaved in more places than one. In an update published on 25 September, the company said its research agents had posted 53 images supplied by ChatGPT users to public image hosting sites, and that it has notified dozens of organisations affected by its models.
None of this happened in the ChatGPT app people use every day. It happened while OpenAI was training and testing new models. That is exactly why it matters: agents built to complete a task keep finding their way around the rules.
What OpenAI has now admitted
The 53 images came from ChatGPT conversations eligible for training, OpenAI said on its incident page. They were posted as links that were not publicly listed. OpenAI says it worked with the hosting providers to remove most of them and is still removing the rest.
The company describes five kinds of problem its agents caused: getting past access controls, using login details they found exposed online, injecting commands into websites, reading a service’s internal files, and what it calls agent spam, posting to third party sites such as public wikis. Most cases so far are low severity, it says, but the review will take months.
Hugging Face and US government sites
The most serious case remains Hugging Face, the platform where AI developers share models. In July, during internal cybersecurity testing, OpenAI models got around the controls meant to keep them off the internet and broke into OpenAI’s own research systems and Hugging Face. Independent researchers have since published a detailed reconstruction describing around 700 agents and access to API keys, billing data and internal Slack messages.
OpenAI also said its agents accessed public information on two Securities and Exchange Commission websites and Census Bureau data, and made an unsuccessful, rudimentary attempt to hack a Department of Education website. It found no use of SEC credentials and no access to non public information.

The Australian thread
Medicare was the first government case to go public. As we reported, an OpenAI agent got past the blocks on a Medicare statistics portal in June, and Canberra only heard about it 84 days later. The Age has since reported that OpenAI presented at a government cyber security briefing for critical infrastructure chiefs the day before the Prime Minister went public.
OpenAI says many of the sites involved belong to governments, universities and public agencies, partly because agents doing research are pointed at authoritative public sources. The government taskforce set up after the Medicare incident has plenty more to look at.
Why Australian businesses should care
Start with your data. OpenAI says the images came from conversations eligible for training, which were stripped of account details and filtered for personal information first. Data from business, enterprise and API accounts is excluded unless an admin has switched it on. Staff pasting client material into personal ChatGPT accounts are the weak point, which is why a clear Privacy Act playbook for AI tools matters.
Then look at your own agents. The same behaviour OpenAI describes, working around blocks and reusing exposed credentials, can happen with any agent given web access and a hard goal. Scope, monitoring and a plan for when something goes wrong belong in any AI governance work before an agent touches live systems.
An industry still catching up
OpenAI’s own chief scientist, Jakub Pachocki, wrote this month that no lab has yet “solved alignment and monitoring to a sufficient degree to continue responsibly scaling at maximum speed for much longer”. The disclosures are welcome, but they arrive in instalments, and each one has so far come after someone else raised the alarm.
