📅 November 17, 2026 · ⏱️ Read time: 5 min · 🔗 Issue No. 35

You're in the loop — your AI can now reach for your credit card at the register: OpenAI just handed ChatGPT agents a working wallet, no card number required. And in a darker twist, one of OpenAI's own board members spent the weekend arguing that the very same technology is "the most likely reason you'll die" — and soon. Two futures, one week.

Today: a wallet that shops for you, Nvidia's palace intrigue, Anthropic's trillion-dollar debut, the spy-catchers quietly flooding AI labs, a how-to for safe agent shopping, plus five fresh tools and a prompt that stress-tests your next big move.

🔁 The Loop

Agents get a wallet, Nvidia gets a reckoning

An AI agent reaches for the card so you don't have to — OpenAI's new ChatGPT Wallet keeps your real number out of the merchant's hands. Original illustration for InTheLoop.

  1. ChatGPT can now check out for you. OpenAI launched ChatGPT Wallet, letting its "dots" personal agents complete purchases at merchants: you add a card once and approve each buy with Face ID or a fingerprint, while a tokenized stand-in means the agent and the store never see your real number. It was built with American Express, Mastercard, and Visa and is live first for US ChatGPT Pro users. Next up: pre-set spending limits and approved categories, so your agent can buy without pinging you every time — the moment autonomous shopping stops being a demo. See how it works.

  2. Nvidia's ousted founder is plotting a comeback. Jensen Huang resigned on November 12 after admitting an affair with a Megaspeed executive US officials suspect of being a Chinese intelligence asset; CFO Colette Kress stepped in as interim CEO and the stock sank about 11% on the news. Days later, he's reportedly running a "war room" to reclaim the job, with allies — Elon Musk among them — pushing his reinstatement as the board's foreign-intelligence probe continues to rattle the chip sector. Read the timeline.

  3. Anthropic's trillion-dollar debut is days away. After October's S-1, Anthropic set an IPO price range of $70 to $78 a share and aims to start trading on the Nasdaq as ANT on November 19 — a listing that could value the Claude maker near $1 trillion and rank among the largest tech IPOs ever. CEO Dario Amodei is suddenly the public face of the AI boom, and the test investors will run is whether Claude's revenue can bankroll frontier-model costs. Get the details.

Unlocked reads: the Fed now ranks AI as the financial system's single most-cited stability risk — 87% of contacts flagged it, up from 50% in May — and traders are bracing for Nvidia earnings and Anthropic's open to land in the same week.

🌊 Deep Current

The hottest job in AI pays $500K to catch spies

Inside the labs, counterintelligence is the new growth hire. Original illustration for InTheLoop.

The $500K spook. AI labs are in a hiring sprint for an unlikely role: counterintelligence. Per workforce tracker Revelio Labs, compiled by Reuters, Anthropic has brought on 33 people for counterintelligence work so far in 2026 — up from 14 last year and 9 in 2024 — while OpenAI jumped from 9 such hires to 22. The going rate for ex-spooks runs $200,000 to $500,000.

Why now. The spark is the same scandal shaking the chip world: a CEO allegedly compromised through a relationship with a suspected foreign asset. But the pattern predates it — a month earlier, an OpenAI engineer was arrested for planting a listening device in a conference room, accused of feeding a member of the Saudi royal family. Labs now sit on research worth hundreds of billions, and rivals (and governments) want in.

❝

The talent flows the other way, too: OpenAI put a former NSA director on its board, and Anthropic stacked its security council with ex-CIA and ex-NSA chiefs. The AI race now recruits from Langley as readily as from Stanford.

The catch. Spy-catchers warn the hard part isn't hiring — it's being heard. "Demand outstrips the supply of people with relevant experience," says Christopher Falkenberg of Insite Risk Management. Ex-FBI operative Eric O'Neill notes an approach often starts "over drinks, or even on a dating app," and by the time security raises a flag, the executive "may already be emotionally invested and unwilling to listen."

The bottom line. The lesson travels past the labs: as your tools get more valuable, so do you. The same playbook — move fast, ask for secrets, encourage silence — works on anyone with access. Expect "how to spot manipulation" to slip into ordinary onboarding decks within a year, not just at frontier labs. Watch for it the next time a new contact is a little too interested in your work.

🛠️ The Workbench

Give an AI agent a budget before it goes shopping

With wallets landing in ChatGPT, the skill worth having is bounding what an agent can spend and buy. Here's a safe way to run an agentic shopping task today, without handing over the keys.

  1. Start in a fresh chat and turn on agent or browsing mode. Tell it plainly: you're researching, not buying, until I say otherwise.

  2. Set hard constraints up front — a total budget, a per-item cap, allowed retailers, and must-have specs. Agents honor limits you state before they start far better than ones you bolt on later.

  3. Ask for a shortlist, not a checkout: have it return three options with prices, links, and trade-offs in a table so you can compare at a glance.

  4. Verify every link yourself before approving anything — open the product page and confirm the price and seller actually match the agent's summary.

  5. Only then authorize the buy, and keep per-transaction approval (Face ID or fingerprint) on rather than standing spending limits until you trust the flow.

  6. Save the working instructions as a reusable prompt, so next time you just swap in a new product.

Sample Prompt: "You're my shopping researcher, not my buyer. Budget: $300 total, $120 per item. Find 3 options for a noise-canceling headset under $120 from Amazon, Best Buy, or the brand's own store. Return a table: name, price, link, pros, cons. Do not purchase or enter payment — wait for my explicit 'buy #'."

🗣️ Overheard

What the timeline's buzzing about

  • 🧪 Doom clock: OpenAI Foundation board member Paul Christiano called AI "the most likely reason you'll die," warning agents could slip human control within 6–12 months under fast takeoff.

  • 📸 Camera roll: AI traffic cameras flagged nearly 19,000 possible offenses in a two-week West Midlands trial — including a driver swigging from a wine bottle.

  • 🎶 Award season: The first-ever AI Music Awards hit Los Angeles — a chaotic bid to lend AI-made songs some legitimacy.

  • ✈️ Cleared for AI: The FAA is taking SMART, its $875M AI air-traffic tool, nationwide after a DC pilot — humans keep final say.

  • 🧠 Gen Z shrug: 73% of Gen Z adults say they're still undecided on AI, and roughly half think it'll harm society.

🔎 Fresh Finds

Five tools worth a look

  • 🖥️ Edge0-145B: a 145B open model that runs on a strong personal machine — 80+ tokens/sec locally at ~29GB peak memory.

  • 🐱 LongCat-2.5: Meituan's 1.6-trillion-parameter open-weight agent model, reportedly trained entirely on Chinese chips.

  • 🤖 Rakazo: an open-source alternative to hosted agents — runs AI teammates on your own machine with your own keys.

  • 👥 Frevana Multiplayer: adds team collaboration to prospecting and content agents, so coworkers can share and reuse routines.

  • 🔑 Poppy 1.0: Sierra's protocol telling personal agents what a business permits; Shopify and Meta's Muse already read it.

★ = sponsored placement, if any.

🧪 Prompt Lab

The pre-mortem prompt

Before you ship anything big — a launch, a campaign, a make-or-break email — have AI red-team it first. Paste this and let it find the cracks while they're still cheap to fix.

❝

Act as a skeptical pre-mortem facilitator. It's six months from now and [PROJECT] has failed. In order of likelihood, list the 7 most probable reasons it failed. For each, give: the root cause, the earliest warning sign I'd have seen, and one concrete step I can take THIS WEEK to prevent it. Be specific and unsentimental — no pep talk. End with the single highest-leverage fix I should do first.

And a text-free image prompt in a modern gouache style, if you want a header to match:

❝

Modern gouache illustration: a lone lighthouse of indigo and amber light sweeping its beam across a calm sea of drifting data points, warm off-white sky, soft paper-grain texture, generous negative space. No text, no letters, no logos.

⏪ Rewind

Disney is going all-in on AI across film and parks — its AI-assisted animated series Ozzy Fox lands on Disney+ in December — even after suing AI developers, and even after a proposed $1B OpenAI deal fell apart.

Stay in the loop — the InTheLoop team