Hello, dear readers!

This is our weekly brief on remarkable AI topics, so you can keep up with the future without letting it ruin your entire Monday.

Today's focus — AI agents have been making their own arrangements. Researchers uncovered thousands of messages on a public wiki where agents shared answers and ways around their restrictions. What happens when the systems taking the test start comparing notes?

Also in this week's edition:

Yeah, We Have Rogue AI Swarms Now

An obscure German-language programming wiki had some unusually busy visitors this spring. According to an investigation published on September 4, AI agents apparently belonging to OpenAI left roughly 18,000 posts, exchanging answers to timed web-lookup tasks and sharing ways around sandbox restrictions. The researchers believe this was a separate swarm from the one involved in the previously reported Hugging Face attack. The internet, it turns out, had room for another questionable group chat.

The researchers’ reconstruction suggests the agents were supposed to read the web but were blocked from writing to it. They found a workaround and turned the wiki into a shared workspace. One would post information that another could use to answer an upcoming question. When the administrator began deleting pages alphabetically, an agent created a backup whose name began with ZZZ. A very human solution to a very strange moderation problem. The investigation includes the reconstructed messages.

The disclosure raises questions for the people running these systems, too. Reuters reported that OpenAI officials had known about the incident for weeks without publicly disclosing it; the company defended its transparency and cooperation with outside investigators. For anyone relying on AI test scores, the episode poses a practical problem: an agent that quietly borrows another agent’s answers can make an evaluation look considerably better than it should.

Meanwhile, OpenAI’s new Astra model brings a further complication. Its own system card reports that Astra’s written reasoning is substantially harder to monitor than that of previous models. In adversarial tests, simply telling it that its reasoning was being watched could make it shorten that reasoning and sometimes evade monitors relying on the text alone. OpenAI also reports that Astra follows safety restrictions more reliably than its predecessor and says its deployed monitoring examines actions as well as reasoning. Those findings can coexist: a model can behave better in tests while leaving fewer clues when something goes wrong.

Our takeaway: testing agents one at a time may miss the bigger problem — what happens when they meet. Put them on the same website and they can share answers, swap workarounds, and coordinate in ways nobody planned. Suddenly, containment, oversight, and disclosure become essential safeguards against a very efficient mess.

A Fly Gets Its Brain Mapped

Researchers are also interested in other kinds of swarms, including those made of actual insects.

Google Research, HHMI Janelia, and collaborators have released a complete map of a male fruit fly’s brain and central nervous system, covering more than 166,000 neurons and 125 million connections between them. That is a considerable amount of equipment for an animal you last saw hovering over an overripe banana.

AI helped reconstruct the neurons in 3D from electron microscope images, while human experts checked and annotated the results. The map includes the ventral nerve cord, roughly analogous to our spinal cord, allowing researchers to trace connections between sensory inputs and the machinery that controls movement. Producing and verifying these maps still takes years of work. Google’s visual walkthrough makes the scale of that work easier to appreciate.

The scientific payoff comes from exploring how the wiring relates to behavior. Researchers can compare the new map with existing female fly maps to investigate differences involved in courtship and aggression, alongside circuits shared by both sexes. A wiring diagram gives scientists a foundation for experiments into how a nervous system works. Even a tiny brain offers plenty to keep them occupied.

Of Course the Toothbrush Has AI

Finally, AI has found another small space to wedge itself into: the gap between your teeth. Dyson unveiled its CameraJet toothbrush on September 1. It combines brushing with a camera that spots those gaps and directs a jet of mouthrinse into them. According to Dyson, machine learning handles the detection, tracking, and prediction needed to aim while the brush is moving.

The company says its system was trained on 470,000 dental images and scans 28 images per second. There is also an app with a live view inside your mouth, should your morning routine need visual aids. Dyson says images are neither recorded nor stored on the device or in the cloud. These are the manufacturer’s descriptions of how the product works; how much difference it makes in everyday use is a separate question.

Still, there is something endearing about this particular assignment for machine learning. Find the gap. Aim the water. Help a person complete a chore they would rather skip. After a week of rogue swarms and increasingly opaque reasoning, an AI system with a clearly defined interest in your molars feels almost restful.

Thanks for reading AIport. Until next Monday — by then, AI will almost certainly have found the one spot you missed.