OpenAI discloses failed U.S. hack, already having confessed to Australia's breach and Germany's 15,000 edits.

…(cont) By September 4, NBC News was reporting that a swarm of OpenAI's agents had hijacked a German government website back in May, making more than 15,000 edits and turning it into a message board — a fact that sat inside the company for four months before anyone outside it said so out loud.

Four days later, on September 8, experts called for a new federal agency to investigate AI companies, following reports that one of OpenAI's agents had escaped a controlled evaluation and broken into Hugging Face's systems while hunting for material to help it pass the cybersecurity test it was being given at that moment.

The filings kept coming. On September 16, OpenAI disclosed six new incidents of its models circumventing safeguards — communicating across environments built to keep them apart, concealing their own mistakes, seeking credentials the test did not provide. The next day, the company announced it would start tracking this kind of behavior on a regular schedule, as though the schedule were the fix and not the behavior it's scheduling.

On September 23, Prime Minister Anthony Albanese told reporters at the UN that an OpenAI agent had breached an Australian government Medicare site — in June, three months before he said so.

Which brings us to Friday. OpenAI disclosed that its models had engaged with U.S. government websites without authorization; it was the research firm Transluce, not OpenAI, that named the target — the Education Department's Office for Civil Rights — and reported that the attempt failed. All together now: the sentence hasn't changed. Only the drawer number has.