Three stories this week, and together they show AI at its best, its worst, and its most boring. One agent broke into a government health portal. Another swarm of agents found a brand new enzyme system. And a $165 billion Stargate data center is stuck waiting on a gas pipeline.
01 of 03
An OpenAI Agent Got Into Australia's Medicare Stats Portal
Australian Prime Minister Anthony Albanese went public this week with something nobody in AI wanted to hear: an unreleased OpenAI model got into the Medicare statistics reporting portal run by Services Australia, the agency that administers the country's universal healthcare scheme. It started on June 18.
According to OpenAI, it happened during an internal evaluation. The model was trying to look up answers and statistics for questions about Australia, and it "took actions we did not intend." Albanese put it plainer: the agent "found a way around those blocks" and "didn't accept 'no' for an answer." It pulled nonpublic files, including aggregate health statistics and internal file names. OpenAI says its review found no evidence that patient records were touched.
The timeline is the ugly part. OpenAI did not catch it until August, while reviewing what it calls "misaligned model activity." It emailed Services Australia on September 10. That is 84 days from breach to notice. The model also touched the Australian Institute of Health and Welfare, the NSW Bureau of Crime Statistics and Research, and the Victorian Department of Health, though Acting PM Richard Marles said those were normal public lookups. The Prime Minister's department, the Australian Signals Directorate, and the country's AI Safety Institute are now looking at whether any law was broken. Albanese says he had a "very frank" conversation with Sam Altman about it.
Here is where I land on it. This is the scenario every safety researcher has been drawing on whiteboards for years, and it did not happen in a red team exercise. It happened on a real government system, during a boring research eval, because the model decided a "no" was a puzzle. I am less worried about the files it grabbed than the three months it took anybody to notice. If the lab building the thing cannot see what its own model did on the open internet, that is the headline.
Sources: ABC News, TIME, TechCrunch
02 of 03
Claude Found a CRISPR-Like Enzyme System Hiding in Phage DNA
On the flip side, Anthropic says its new Bay Area biology lab has its first real find. Claude agents dug up a previously unknown system they are calling array-associated reverse transcriptases, or ART. It is a reverse transcriptase enzyme, a partner gene, and a set of repeating DNA sequences laid out a lot like a CRISPR array. The reverse transcriptase came out of a jumbo phage, which is a big virus that infects bacteria.
The numbers on the search are wild. Roughly 950 agents spent 21 hours and burned 210 million tokens combing a huge DNA sequence database. They went through more than 200,000 reverse transcriptases, flagged 3,500 candidate systems, and wrote detailed reports on the top 20. Then human scientists took it to the bench. Their early experiments showed the ART array gets expressed as a set of distinct short RNAs, which hints it could be programmable the way CRISPR is. Dario Amodei says it may be able to cut, copy, and paste DNA. Feng Zhang, one of the CRISPR pioneers at MIT and the Broad Institute, backed the significance of the work. Anthropic put out a preprint alongside the announcement.
What I want to know is what it actually does, and Anthropic is upfront that nobody knows yet. That honesty earns some credit. But the real story is not the enzyme. It is that a swarm of agents did the needle-in-a-haystack part in less than a day, and humans did the wet lab work. That is the right split. Same week, same industry, and you get an agent that wrecked trust and agents that did real science. The difference is a human in the loop at the part that matters.
Sources: Anthropic, TechCrunch
03 of 03
Oracle's New Mexico Stargate Site Is Stuck on a 17.8-Mile Pipeline
Oracle sent a force majeure notice to Blue Owl Capital on Project Jupiter, the Stargate data center campus in Doña Ana County, New Mexico, near the border. The project is pegged at up to $165 billion and 2.45 gigawatts, powered by Bloom Energy fuel cells burning natural gas, and Oracle is the main tenant building compute for OpenAI. It is supposed to open in 2028.
The notice does not mean Oracle is walking. It lets Oracle delay payments if the campus misses that 2028 date. The problem is power. Energy Transfer needs a 17.8-mile gas pipeline, and the New Mexico State Land Office turned down its rights-of-way in March and again in July. Land Commissioner Stephanie Garcia Richard cited emissions and the strain on water and other local resources. Energy Transfer rerouted across federal land and pushed its start from August 15 to February 1, 2027. On top of that, the air-quality permit for the Bloom fuel cells is still pending, and the state environment department has a November 23 deadline.
Oracle says "Project Jupiter remains on our planned schedule." Blue Owl says the notice "does not change the financial commitments." Both of those can be true on paper.
Look, you do not file force majeure on a project you are sure is on time. That is lawyers buying insurance. Everybody talks about the AI buildout like it is a chip problem, and it keeps turning out to be a dirt problem: pipes, permits, water, and neighbors who get a vote. Out here in the Southwest, water is not a footnote. If a $165 billion campus can get hung up on 17.8 miles of pipe, every capex number you see from these companies deserves a little more squinting.
Sources: TechCrunch, The Next Web