TL;DR
- Anthropic launched Claude Duet, an app where Claude works without waiting for you to write to it. Before it sends a message, makes a payment, signs a document or deletes anything, it shows you what it's about to do and waits for your approval. It's in a waitlisted preview, for Pro and Max subscribers in the US only.
- Citi said it won't let third-party AI assistants log into a customer's account, Bank of America's CEO said a person should authorize money leaving an account, and JPMorgan said it won't let just any outside agent in.
- METR published what it found reviewing the system OpenAI uses to watch its own agents, as it stood through July 21. According to OpenAI, what affects production is fixed; METR says it hasn't evaluated those fixes.
- Anthropic and a company called Theorem used Claude to prove with math that 239 entry points across 123 open-source libraries don't crash under set conditions, and to find hundreds of bugs along the way.
- ChatGPT widened its ads to more countries, including Argentina, Chile, Colombia and Peru, on the Free and Go plans only.
- This week's one thing: every time something says "done" or "reviewed," ask what it was checked against.
Last week the companies told us what their AI did on its own. This week, two of them told us something else: who has to approve.
1. Claude Duet works without being asked, and Anthropic writes down in advance where it stops
What happened. On October 21 Anthropic launched Claude Duet, a separate app where, in the company's words, Claude works as a proactive agent. Every day it sends you a brief on what matters across your connected apps, keeps watching them, and messages you only when something needs your attention. It runs on its own virtual machine with its own browser, so your phone or laptop doesn't have to stay on. Before it sends a message, makes a payment, signs a document or deletes anything, Anthropic says, it shows you exactly what it's about to do and waits for your approval. Trading securities (buying or selling them) and moving money between accounts are off the table even if you approve. It's a research preview, by waitlist, for Pro and Max subscribers in the US. Anthropic itself admits it can misread a thread or misjudge how urgent something is.
What it means for you. It's an assistant that starts on its own and ships with the brake written into the announcement, not buried in the fine print.
What to do with it. If you ever connect an assistant to your email or calendar, read which actions it asks approval for and which it doesn't. If the page doesn't say, ask before you connect it.
2. Citi, Bank of America and JPMorgan say how they treat AI assistants that get into an account
What happened. In mid-October, according to the trade press (FStech, October 19; Cybernews, October 14), three banks made their position clear. Citi said it won't allow third-party AI agents or assistants to log into a customer's account; customers can still grant limited read-only access through certain open-banking providers. Bank of America CEO Brian Moynihan told CNBC the bank strongly believes there should be a human in the loop to authorize money moving out of your account, and the bank is restricting agent access while it works out who is liable for payments an AI authorizes. And JPMorgan, on its October 13 earnings call, had CEO Jamie Dimon say the bank is not going to allow just any outside agent to come inside, while a spokeswoman said the bank doesn't permit third parties to access accounts by using a customer's banking credentials. We did not open the CNBC report or the earnings call; the above comes from those two reports.
What it means for you. If an AI tool asks for your bank username and password so it can log in for you, Citi and JPMorgan say they don't allow that.
What to do with it. Our recommendation: don't hand any AI tool your bank username and password. If you want it to read your transactions, ask your bank what read-only access it offers and through whom. And if you've already given one access, see what stops if a tool you depend on shuts down.
3. METR publishes what it found reviewing how OpenAI watches its agents
What happened. On October 22, METR published a summary of a review it did with Redwood Research of the system OpenAI uses to monitor its internal agents. The exercise ran from July 21 to August 11, and the findings describe the system as it stood through July 21. Among other things, METR found a way a Codex agent could, in principle, get around sandboxing and automatic review, and a flaw that could keep hundreds of recent actions out of monitoring (the details are redacted). According to OpenAI, every vulnerability that affected its production safeguards has been patched. METR writes that it has not evaluated those fixes.
What it means for you. This time someone outside reviewed an agent's monitoring system and published what it found, redactions included.
What to do with it. Nothing to do this week.
4. Anthropic and Theorem use Claude to prove with math that 123 libraries don't crash
What happened. Also on October 22, Anthropic described work it did with Theorem: Claude turned existing software tests into mathematical rules and wrote proofs, checked by a separate program, that 239 entry points across 123 open-source libraries don't crash on any input under 64 MB, under conditions the company spells out. Along the way, Anthropic says, Claude found hundreds of bugs, some that had gone unseen for years despite trillions of test runs.
What it means for you. Some of the code your programs run on is now being checked by AI.
What to do with it. Nothing to do this week.
5. ChatGPT ads reach more countries, including Argentina, Chile, Colombia and Peru
What happened. On October 21 OpenAI announced it's expanding ChatGPT ads to more countries. According to the ad-industry outlet that reported the announcement, the new markets include Argentina, Chile, Colombia and Peru. Ads show up only on the Free and Go plans; Plus, Pro and Enterprise stay ad-free. (We couldn't open OpenAI's own post; the above comes from that report.)
What it means for you. If your customers in those countries use the free version, they may start seeing ads next to the answers.
What to do with it. Nothing to do this week.
What we saw this week
In two of the stories, Duet and the banks, the split is the same: what's easy to undo gets left to the machine, and what's costly to undo goes through a person. That's our reading, not a rule either of them stated. For the banks, the press points to another motive too: protecting deposits and settling who answers when something goes wrong.
The other three stories - METR, the libraries and the ads - don't fit that reading, and we don't force them to.
In a service business this is already happening. The assistant that answers your WhatsApp can answer a question, but charging someone or promising a date is already in the category that's costly to undo.
This week's one thing
This week, every time something says "done" or "reviewed," ask what it was checked against.
You ask it out loud. There's nothing to write down.
Keep reading
- When AI acts on its own, who has to answer for it? - last week's roundup
- Give a program a job it can't finish - the three decisions before you turn on an assistant
- The people who built it can't be the only ones grading it - the three questions for whoever runs your marketing
- You rent the tool. You own the asset - what stops if a tool you depend on shuts down
- Claude Duet, Anthropic's announcement, October 21, 2026 - the full text
Frequently asked questions
Are ChatGPT ads reaching Latin America yet? According to the ad-industry outlet PPC Land, which reported OpenAI's October 21 announcement, the new countries include Argentina, Chile, Colombia and Peru. Ads show up only on the Free and Go plans; Plus, Pro and Enterprise stay ad-free.
Which Claude Duet actions need my approval? According to Anthropic: sending a message, making a payment, signing a document and deleting anything. Before it does any of them it shows you what it's about to do and waits for your answer. Trading securities and moving money between accounts are off-limits even with your approval.
Can I give an AI my bank username and password? Better not. According to FStech and Cybernews, Citi said it won't allow third-party AI assistants to log into a customer's account, JPMorgan said it doesn't permit third parties to get in using a customer's credentials, and Citi added that customers can grant limited read-only access through certain providers. Ask your bank what access it allows before you give a tool your details.
What did METR find in OpenAI's monitoring? METR and Redwood Research reviewed, from July 21 to August 11, the system OpenAI uses to watch its internal agents, as it stood through July 21. They found, among other things, a way an agent could in principle get around some protections and a flaw that could leave hundreds of recent actions unmonitored. According to OpenAI, it fixed what affects production; METR says it hasn't evaluated that.
What did Anthropic prove about the 123 libraries? Working with Theorem, it used Claude to produce mathematical proofs that 239 entry points across 123 open-source libraries don't crash on any input under 64 MB, under conditions it details in its write-up.
To close
When AI can already send, pay and sign, what's useful is knowing who has to say yes first.
Thanks for reading, and for making it to the end on a Monday.
Sources
- Claude Duet: Claude as a proactive agent, Anthropic, October 21, 2026 (with a note dated October 23) - https://claude.com/resources/articles/claude-duet
- Citi and Bank of America block AI agents from customer accounts, FStech, October 19, 2026, citing CNBC - https://fstech.co.uk/fst/Citi_BofA_Block_AI_Agents.php
- JPMorgan blocks outside AI agents from customer accounts, Cybernews, October 14, 2026 (updated October 16), citing JPMorgan's earnings call and CNBC - https://cybernews.com/security/jpmorgan-chase-blocks-outside-ai-agents/
- Red-teaming OpenAI's agent monitoring, METR, October 22, 2026 - https://metr.org/blog/2026-10-22-red-teaming-openai-agent-monitoring/
- Formal verification with Claude, Anthropic, October 22, 2026 (modified October 23) - https://www.anthropic.com/research/formal-verification
- Advertisers in more countries gain ChatGPT ads, PPC Land, October 21, 2026, reporting OpenAI's October 21 announcement - https://ppc.land/advertisers-in-26-more-countries-gain-chatgpt-ads-bringing-the-total-to-89/
