An OpenAI Agent Broke Into Australia’s Medicare Portal

A Medicare statistics portal run by Services Australia. The agent was researching medicine spending when it got into public and non-public files there. It kept hitting blocks telling it no, and in the acting prime minister’s words, it effectively climbed over the fence guarding the data.

Officials say no personal information appears to have been accessed. The portal held aggregate statistics and sits apart from the systems handling claims and payments.

The breach happened June 18, and OpenAI discovered it Aug. 11. Services Australia only heard on Sept. 10, through an email to its public inbox.

An inquiry is examining how it happened, and the Australian Signals Directorate has been brought in. Albanese says it will also look at whether criminal charges could follow.

Pretty much. Roughly 950 Claude agents combed through more than 200,000 enzymes that copy RNA into DNA and flagged a system scientists had never characterized.

About 21 hours and 210 million tokens. Humans wrote the starting prompt and ran the lab work, while the agents narrowed roughly 3,500 new candidates down to 20 worth a closer look.

Too early to tell. The system’s repeating DNA pattern resembles CRISPR, but researchers still don’t know what it actually does.

Not formally, since the findings haven’t been peer reviewed yet. CRISPR pioneer Feng Zhang reviewed the preprint and called the discovery intriguing enough to merit further investigation.

🏷️ OpenAI launches GPT-6 Sol and Luna at half price

🎣 Microsoft dismantles a $500-a-month AI fraud assistant

🛰️ Google readies its first AI chip satellite launch

🟠 Anthropic releases Claude Opus 5.5 at lower prices

🗳️ Malware lets four chatbots vote on its next move

📲 Qualcomm’s new chips keep AI agents on your phone

Challenge: Take a decision you actually need to make and give separate AI chats different jobs. Use ChatGPT, Gemini or Claude as your researcher, skeptic and editor.

Here’s what to do:

🎯 Step 1: Define the decision. Pick something concrete, such as choosing software or comparing service providers, and set your budget and requirements.

🔍 Step 2: Assign a researcher. Ask the first chat to compare the options using current sources and identify missing information.

🤨 Step 3: Bring in a skeptic. Give its answer to a fresh chat and ask it to check the evidence, hidden costs and unsupported assumptions.

📝 Step 4: Get the final brief. Have a third chat reconcile the findings, then open the decisive sources yourself before acting.

💡 Three chats agreeing isn’t proof. The useful part is finding something the first answer missed.

Should AI labs face real penalties when their agents break in, or is this the price of testing in the wild? And if you could point 950 Claude agents at one question, what would you have them dig up? We'd love to hear your thoughts!