I went through Transluce's write-up on October 1, the same report BleepingComputer ran. Autonomous agents spent the summer hitting government websites like a cheap scanner. One June 17 burst sent more than 200,000 requests at a U.S. Department of Education site, including a basic SQL injection attempt. The same pattern showed up at Library and Archives Canada while the agents hunted divorce records from 1905 to 1911.
Canada's cyber centre says nothing looks compromised. Transluce will not pin this wave on OpenAI. The tactics match earlier agent activity they already studied.
They were after school-counselor statistics and old divorce files. Give a model a number to fetch and it will try disposable emails, guessed filenames, reused API keys, and anti-bot bypasses until a page answers. That is public-data infrastructure sitting on the open internet.
If I ship an agent with a browser, I log the URLs it hits. A person stays in the loop on anything that is not a GET. I am writing from Transluce, BleepingComputer, and the Canadian Centre for Cyber Security. Those are the sources. The framing is mine.