iPrompt
THE AI NEWSLETTER THAT TURNS NEWS INTO ACTION
ISSUE #162 · WEDNESDAY · 18 NOVEMBER 2026
Stop feeding the feed.
You’ve put in the work to build an audience. But when the only way to reach your fans is through a social feed determined by algorithms, the platform decides who sees what you make next.
beehiiv gives you the tools to reach your audience directly. On their all-in-one platform, you can turn followers into subscribers, grow your newsletter, and connect it all to your community or podcast. Stop chasing virality on platforms you can’t control. Build a stronger, more sustainable business on beehiiv.
THE HOOK
On Friday, OpenAI gave its agents a wallet. Its terms already say whose purchases they are: ‘You are responsible for transactions made on your behalf.’ United and Railway faced the same question this week. One locked agents out. The other is covering a bill run up on stolen keys. Last week we listed what a task contains. This week: what it can spend.
OUR ANGLE
The agent can spend. The bill has your name on it.
The wallet. ChatGPT Wallet, rolling out since Friday to US Pro subscribers, lets OpenAI’s ‘dots’ agents pay with a stand-in card number once you approve each purchase with Face ID or a fingerprint. OpenAI plans limits by amount, category and time period, so agents can buy without the tap. Section 14 of its service terms already puts what the agents do, and buy, on you.
The bouncer. United’s answer, on 11 November: unauthorised agents may not search, book or manage trips on its site, its app or through agencies, even with the customer’s permission. Its reasons include unintended purchases and check-in steps a human must confirm. Its customer guidance says it doesn’t yet work with any third-party agent.
The breach. Railway’s incident report, published on 10 November, says the attackers in its October breach prioritised AI provider keys stored by customers. Spend on the stolen keys reached an estimated $740,000. Refunds depended on each provider’s policy. One enforced a no-refund rule, so Railway is paying that share itself.
My read. This part is interpretation, not legal advice. ‘Who pays?’ has three answers. A purchase you approve is yours. An agent buying the wrong thing within the authority you gave it is, under OpenAI’s terms, also yours, though returns and card protections may help. Spending on stolen credentials falls to fraud rules and refund policies, which at Railway varied by provider. In all three, a limit decides how big the loss can get. The prompt below finds what can spend.
A fictional nine-person agency maps the eleven things its AI tools could spend on, item by item: the setting before, the setting after, and which limits it actually tested. Plus Visa’s test of 22 AI shoppers. Read the companion.
Prediction: by 30 June 2027, OpenAI will let at least some Wallet users pre-approve purchases within a spending limit, and its terms will still say you’re responsible for transactions made on your behalf. Autonomy will arrive before liability moves. The companion has the counting rules.
OpenAI · Wallet · OpenAI · service terms · OpenAI · Wallet terms · Skift · United · Unite.AI · United’s guidance · Railway · incident report
PROMPT OF THE WEEK
Map what your AI can spend
Run this for yourself, then for your team. It interviews you and builds one table. Allow 15 minutes, with last month’s card statement and your AI billing pages open.
Help me find everything connected to AI that can spend my money, or my company’s, and what would stop each one. Don’t recommend products. Never ask for a key, password or card number. If I paste one, tell me to delete it and rotate it.
SCOPE: [me / my team of N / my company]
WHAT I KNOW WE USE: [AI apps, agents, automations, API keys, or ‘not sure’]
1. Interview me one question at a time. Walk me through card statements, AI billing pages, automation tools, browser extensions, agents with a saved card or login, and keys given to freelancers or apps.
2. Stop after twelve questions, or sooner if nothing new comes up.
3. For each item record its CEILING as STOPS, ALERTS ONLY, NONE or UNKNOWN. Keep UNKNOWN as UNKNOWN, but rank it as if it had no ceiling.
4. WORST DAY: the most it could spend in 24 hours if it misbehaved or the credential were stolen. Use only figures I give you: the fastest spend I’ve seen, any usage or rate limit, any remaining balance or cap. Show the arithmetic. Without those figures, write NOT ESTIMATED.
5. For the top three, give one change each. It must include a hard limit or removing access. A separate key or card only counts if it has its own limit.
6. Output one table: WHAT, WHERE THE CREDENTIAL LIVES, WHO CAN USE IT, NORMAL MONTH, CEILING, WORST DAY (with arithmetic), WHO PAYS IF MISUSED (as I understand it, or UNKNOWN), CHANGE, VERIFY. Under VERIFY, list what I must check myself in the provider’s settings, card terms or contract.
Why it works: ask ‘are we exposed?’ and you get a shrug. Walk through the statements and you find the key a freelancer still has. The VERIFY column matters most: an interview can’t confirm a setting or who pays. Only the settings page and the terms can.
ALSO THIS WEEK
Google’s new fast model reaches free users
Gemini 4 Hydrogen arrived on 12 November. Google says it finishes tasks 2.5 times faster than Gemini 3.8 Flash and cuts fabricated claims more than fivefold. Free users get it in the Gemini app, when auto selection picks it, and in Search. Google’s figures: test it on your own work.
An open-source attack agent walked in with valid keys
Arctic Wolf linked AntiFox, an AI penetration-testing tool anyone can download from GitHub, to intrusions that reached production secrets in AWS. In some cases the attackers already held working credentials, and Arctic Wolf doesn’t know how they got them. The Cloud Security Alliance’s verdict: long-lived credentials remain the core weakness.
Apple wants 200,000 people to test health AI
Apple and OpenEvidence opened the Apple Health AI Usage Study on 15 November, for US adults aged 21 and over with a recent iPhone and Apple Intelligence switched on. Answers cite clinical sources such as NEJM and JAMA. It’s a study of trust and understanding, not yet a verdict on accuracy.
TOOL OF THE WEEK
TruffleHog
TruffleHog is a free, open-source scanner that searches folders, git history and other places for keys and passwords, then tests whether each one still works. It recognises more than 800 kinds of credential.
Where I’d point it: any repository or old project folder where someone may have pasted an AI key. Results come as verified, unverified or unknown. Rotate verified keys today. Unverified isn’t proof a key is dead, so rotate any that matter.
The catches. Testing means it tries each key with the provider, so scan only what you own. It runs from the command line, so most teams will want a developer. It won’t find keys taken from inside a platform, as at Railway. And I haven’t run it on a live system: this rests on its documentation and reviews.
Verdict: worth one afternoon on the folders your AI projects live in.
TIP OF THE WEEK
Check that your limit stops something
For every AI account you pay for by use, open the billing settings and find the limit on each project or workspace. Then check what it does when it’s reached. On OpenAI’s platform, a spend alert sends a notification and traffic continues. Requests only stop if you switch on ‘Enforce a hard limit’. Anthropic’s API runs on prepaid credit. Auto-reload, if it’s on, tops the balance up whenever it runs low, so set a spend limit as well.
The limits. Enforcement isn’t instant: OpenAI says recorded spend can slightly exceed the limit. A hard stop takes your own app down too, so my starting point is about twice a normal month, adjusted for what an outage would cost. And none of this covers a card saved in an agent. That has its own limit.
Pro move: one project or workspace per job, each with its own key and its own limit. The key lets you see and revoke one job’s spending. The limit is what caps it.
YOUR MOVE
Put a ceiling on one thing that can spend
Run the prompt. Take the item with the worst plausible day and fix that one before Friday: a hard limit, or remove its access. If it isn’t yours to change, send the table to whoever owns it.
Then trip the limit once, on a spare project with a tiny cap, so the first time it stops isn’t in front of a customer.
REPLY WITH ONE LINE
What did you find that could spend, and what stops it? Reply before you’ve fixed it if you like. For example: ‘Zapier’s AI key: no limit, auto-reload on. Asking our developer for a hard stop.’ Never paste a key. I read every reply.
Know someone whose agent has a saved card? Forward them this first.
Attio is the #1 fastest-growing CRM on Brex and Ramp.
Built for the most ambitious companies in GTM, Attio is an end to end revenue platform that allows teams to run multiple plays across different revenue motions and lifecycles.
Teams at the frontier are already building on it: Granola, turbopuffer, Modal, and many many more.


