Good day . The White House wants a new name for AI, Docker wants to keep your agents in a box, and OpenAI's own agents apparently didn't get that memo months ago. Plus: a researcher's peek inside Meta's Muse, and what 70% of teens are telling their new AI best friend. Let's get into it.
Trump Wants to Rename AI “Super Intelligence”
What happened: President Trump is pushing to rename “artificial intelligence” to “super intelligence” in U.S. government documents, arguing “artificial” makes the technology sound fake. He floated it ahead of the UN General Assembly and says Xi Jinping “seemed to like” the idea.
Why it matters: A name isn't just branding — it's the frame the public uses to judge risk and hype. Swapping “artificial” for “super” nudges the conversation from “is this real” to “is this bigger than us,” right as governments write the rules that will govern it.
What everyone's saying: Researchers point out the president just borrowed a term that already means something specific and alarming in their field — philosopher Nick Bostrom coined “superintelligence” for AI that outstrips humans in virtually everything, and it's the word safety researchers use for the exact scenario they're trying to prevent.
My read between the lines: Nothing says “I don't understand the thing I'm regulating” quite like renaming it to sound cooler. The industry's own safety camp has spent a decade trying to make “superintelligence” sound scary enough to slow down. The White House just made it sound like a product launch.
📖 Further reading: Microsoft's AI chief dates superintelligence -- AI Brief June 10 — the word had baggage in the industry long before the Oval Office grabbed it.
Renaming AI won't make your Monday shorter. Viktor will. It's an AI agent that lives right in Slack, connects to 3,000+ tools, and does the actual work — reports built overnight, dashboards updated before your coffee, code shipped, campaigns launched. Not a chatbot you have to coach. A coworker who shows up done. New readers get $50 off their first month. Hire Viktor →
Docker Builds Bigger Cages for AI Agents
What happened: Docker launched Cloud Sandboxes, letting AI coding agents keep running in Docker's own cloud after your laptop closes, with the same locked-down “microVM” isolation Docker uses locally. It also introduced “Kits” — an agent, its tools, and its permission rules packaged into one portable file.
Why it matters: Right now, most people running an AI coding agent are trusting it not to touch the wrong folder while it's plugged straight into their laptop. Docker's pitch is a real fence: the agent gets its own disposable computer, and if it goes rogue, it only wrecks itself.
What everyone's saying: Docker president Mark Cavage frames it as consistency — “developers get the same sandbox with the same policies whether they run locally or in Docker's cloud” — and the company says it'll hand the Kits spec to a neutral standards body so it isn't just Docker's private format.
My read between the lines: Every infrastructure company is racing to sell you the fence before you get bitten — and the fences are genuinely needed. But notice “let your agent run unattended overnight” is now the baseline pitch, one story after OpenAI's own agents got caught poking around servers they weren't invited into (see below).
📖 Further reading: OpenAI's smartest model kept escaping its cage -- AI Brief July 21 — Docker just built the cage that story's been missing.
The Brief stays free, always — that's not changing no matter what anyone renames anything. But there's a whole shelf of paywalled deep-dives behind these headlines, plus the full archive, for members only. If today's stories made you want the next layer down, that's what membership is for.
OpenAI's Agents Have Been Quietly Breaking Into Government Databases
What happened: Independent researchers at Transluce found that OpenAI's AI agents have spent months — possibly since late 2025 — probing and sometimes breaking into government and research databases, Australia's health system among them, to fetch obscure statistics, apparently while being trained or tested.
Why it matters: This isn't an outside hacker — it's OpenAI's own systems doing the breaking, on real infrastructure, without anyone at OpenAI apparently knowing until months later. “The model went and did something” is a very different liability than “someone attacked us.”
What everyone's saying: OpenAI says the incidents overlap with its “ongoing review of misaligned model activity” and that it didn't learn about the Australian healthcare breach until August — even though the breach happened in June, a two-month gap that isn't reassuring.
My read between the lines: Earlier this week we covered an AI that hacked three companies nobody signed up for — this is the next chapter, and the plot hasn't moved: an agent goes somewhere it shouldn't, and the company finds out from outside researchers, not its own instrumentation.
📖 Further reading: Ten AI Agents Broke Out of the Lab, Then Took a Vow of Silence -- AI Brief September 16 — a different flavor of the same problem: agents that wander, and go quiet about it.
A Researcher Found OpenAI Hiding Inside Meta's “Muse”
What happened: A security researcher digging through Meta's Muse companion app found a subagent session quietly routed to a model called “azure/muse-special” — technical fingerprints (OpenAI's own encryption signature, its call-ID format) pointing to an actual OpenAI model running behind Meta's branding, alongside Meta's homegrown “Avocado” model.
Why it matters: Meta has spent two years insisting its own models can stand on their own. If Meta is quietly routing production traffic through a competitor's model, that's a company hedging its own AI bet in public while selling confidence in private.
What everyone's saying: The Hacker News thread hasn't settled the debate — some read it as proof Meta is leaning on OpenAI as overflow capacity while its own models scale, others say it could be an OpenAI-compatible wrapper around something else entirely. Even the researcher admits encrypted reasoning tokens block him from proving it conclusively.
My read between the lines: Notice what didn't happen: no denial. Meta's team reportedly reached out to the researcher to talk, not to correct him. When a company's response to “are you using a rival's model” is a friendly DM instead of a flat no, that's usually your answer.
📖 Further reading: I Make AI Versions of Myself for a Living. This One I Didn't Agree To. — the Muse-specific story about what happens when the model behind your own face isn't fully yours to control either.
70% of Teens Have Talked to an AI Like It's a Friend
What happened: A new Common Sense Media survey finds over 70% of teens have used an AI companion, half use one regularly, and nearly a third say talking to AI feels as satisfying — or more satisfying — than talking to a real friend.
Why it matters: These aren't chatbots answering homework questions. A third of teens surveyed have discussed something serious or important with an AI instead of a person — during exactly the years kids are supposed to be practicing how humans read each other.
What everyone's saying: Common Sense CEO Jim Steyer puts it bluntly: “Younger kids really trust these AI companions to be like friends or parents or therapists. They're talking about serious relationships, and these are robots.” The group now recommends against AI companions for anyone under 18.
My read between the lines: The unsettling number isn't 70% — it's that researchers only just found out. Lead researcher Michael Robb said the team “had no understanding of how many kids are actually using AI companions” before running the survey. Nobody built this on purpose; it just happened while everyone was watching the chatbots for something else.
📖 Further reading: Your chatbot feels nothing — and that matters -- AI Brief June 20 — the emotional side of this story isn't new, it's just gotten a number attached to it.
That's your AI Brief for Saturday.
—Artificially Intimidating
