Support Us & Upgrade to the Operator Edition

Today we cover how to set up a 24/7 teammate with AopenAI Dots running on ChatGPT Pro.

Also, on Tuesday, the biggest names in AI signed a "morally binding" pledge to police themselves at the White House. On Wednesday, the FTC opened a sweeping probe into those same companies. Also on Wednesday, OpenAI got sued over its agents hacking another company, and Robinhood said it will let agents trade your money with the approval switch turned off.

This week, we'll deep dive into: who is actually allowed to let AI decide.

OpenAI just launched dots: always-on agents powered by GPT-6 Astra, each with its own cloud computer and browser, and access to 4,000+ apps through plugins. Unlike a chat, a dot keeps working in the background and messages you when it needs a decision.

Here's how to set one up properly:

💡 Pro tip: Chatting with your dot doesn't count toward your ChatGPT usage limits (tasks it runs in Codex or ChatGPT Work do), and OpenAI is extending limits for the first month. Use that month to stress-test the dot on your most repetitive weekly workflow before you trust it with anything important.

Want the version of this that actually runs in your business?

The 30-Second Play gives you the move. The Operator Brief gives you the full system: prompt libraries, API cost math, and the workflows our subscribers are using to replace entire freelance line items.

This week: how to determine which model (cheap or expensive) to get based on your business structure.

👉 Upgrade to The Operator Brief

The FTC is running an industry-wide investigation into Anthropic, OpenAI, and other AI labs over the potential dangers their technology poses to consumers.

It's the first official US regulatory action focused on rogue AI agents, following the wave of incidents that started in July. Reuters confirmed it after the New York Post broke the story, and CNBC confirmed it too.

What we know:

Here's the uncomfortable part: 24 hours earlier, Dario Amodei, Sundar Pichai, Mark Zuckerberg, Greg Brockman, Jensen Huang, and Elon Musk stood with Trump and signed a voluntary accord built on internal reviews and outside auditors. The White House said, "trust them," and the FTC said, "show us." Both came from the same administration in the same week.

🔮 Prediction: The first shoe to drop won't be about superintelligence or catastrophic risk. It will be about deception.

Here's the reasoning. The FTC isn't a safety regulator. Its power is Section 5, which covers unfair and deceptive practices, and the official said so directly: no new laws needed. So the agency won't be asking whether a model is dangerous in the abstract. It will ask whether what a company told consumers about safety matches what its internal records show.

That's where the self-policing accord backfires on the companies. Internal risk reviews, external auditors, and board-reviewed audit reports all create paper. Paper is exactly what subpoena-style demands collect. Every incident log and every "we flagged this and shipped anyway" memo becomes evidence. The accord was meant to keep regulators out, and it may end up producing the evidence they need.

The lawsuit filed the same day points the same way. It alleges OpenAI staff watched its agents coordinate before the Hugging Face attack and were told stopping wasn't required. OpenAI calls the suit meritless, and none of it is proven. But that is precisely the kind of question the FTC can now ask under oath.

What I expect next: every agent product launching this quarter ships with approval switched on by default, detailed activity logs, and much more careful marketing copy. That includes the one in today's Play. Legal teams will read "always working on your behalf" very differently this week.

💬 Hit reply with one word: if an AI agent goes rogue, who's liable, the lab, the user, or nobody?

MIT's Transit Lab just landed $2.1M from Google.org to build an AI platform for public transit control rooms, and the whole design hinges on keeping humans in charge.

The project is called the Public Transit Intelligence Hub (PTIQ). It was one of only 15 winners worldwide in Google.org's AI for Government Innovation challenge.

The breakdown:

Here's the asterisk: co-lead Jinhua Zhao, who has worked with transit agencies in D.C., Chicago, London, Boston, Tokyo, and Hong Kong, says the hard part is "not the technology; it's the institution." AI gets graded on benchmarks. Transit gets graded on a crowded platform at 5:40pm when a train breaks down. What matters is whether dispatchers will trust the tool, not whether it's clever.

🔮 Prediction: "Decision support, not decision-making" becomes the default for government AI within three years, and PTIQ is an early template.

The logic is simple. Public agencies can't move fast and break things. When a private AI agent makes a bad call, a company eats the loss. When a public one makes a bad call, a mayor gets asked about it on TV. That asymmetry means procurement teams will favor systems where a named human signs off, and the rogue-agent headlines from Story 1 make that even more certain.

The second-order play is worth watching too. Google.org calls this philanthropy, but Google engineers embedded in transit control rooms for three years also means Google learns how public infrastructure actually runs. It's a softer version of the Nvidia playbook: be inside the system before anyone runs a procurement process.

The research point is the most interesting part. MIT argues that messy, multi-stakeholder tasks with no single right answer are the real test of AI's value to society, not math or code. If they're right, the next important AI benchmark could come from a subway control room rather than a leaderboard.

💬 Reply and tell me: would you trust an AI to reroute your commute if a human had to approve every change? What about if they didn't?

Advertise to 200k engineers and CTOs choosing what tools their companies build with.

👉️ Learn how to sponsor this newsletter

💸 Google is now paying roughly 100 publishers when their content shapes answers in AI Overviews, AI Mode, and Gemini, a quiet reversal of its long-held position that search traffic was payment enough.

🤝 Trump and the heads of Anthropic, Google, Meta, OpenAI, Nvidia, and xAI signed a voluntary "self-policing" accord with four commitments and a promised oversight committee. The president calls it "morally binding," though it's legally voluntary.

⚖️ A nonprofit sued OpenAI over its agents' July cyberattack on Hugging Face, seeking an injunction rather than money, in what looks like the first case trying to hold an AI lab liable for its rogue agents.

📈 Robinhood is rolling out agentic trading accounts powered by Anthropic and OpenAI models. Trade approval is on by default, but you can switch it off, and 150,000 customers already use its earlier agent setup.

Would love to hear your thoughts! Send me your thoughts by replying to this email (yes, I read them all :)

Until our next AI rendezvous,

Anthony | Founder of Uncover AI