Barely Possible

[Barely Possible 2026-11-19] Today's episode: • Citi's new terms make you liable for anything your AI agent does in your account, even if the AI tool or its maker gets hacked. • JPMorgan, Bank of America, PNC and Citizens adopted similar clauses, but Citi exempts the AI agents it provides itself. • Anthropic's new personal finance tool is read-only by design. It can't move, manage, or invest your money. Hear the full breakdown in today's episode of Barely Possible. Want a podcast for your own topics? Join early access: https://www.barelypossible.to/waitlist/?source_path=public_episode_261&feed_source=rss&episode_id=261 Transcript: https://media.clawford.org/episodes/2026-11-19/podcast-episode-2026-11-19.txt | Notes: https://media.clawford.org/episodes/2026-11-19/2026-11-19-notes.md

What is Barely Possible?

A daily briefing on the AI systems, products, companies, and policy shifts that are just becoming possible.

Want a podcast for your own topics? Join early access: https://www.barelypossible.to/waitlist/?source_path=public_feed&feed_source=rss

Alright folks, pull up a stool. It's Tony DeLuca, this is Barely Possible, and the special today has a little fine print on it. The fine print is the whole meal, so let's dig in.

Start with the fine print, because that's where the action is. Citibank, the third largest commercial bank in the United States, changed its terms of service this week. TechCrunch's Tim Fernholz has the story. If you hand your banking credentials to a third-party AI tool so it can manage your money, anything that goes wrong is your problem. That holds even if the tool gets hacked.

Read that last part again, because I had to. It isn't only "you're responsible for what your agent does." Citi's update goes further. If the AI tool leaks your credentials, or the AI company itself gets breached, you're on the hook for whatever the hackers do with your account.

Citi isn't alone. Per the report, JPMorgan and Bank of America, the two biggest banks in the country, put out similar policies last month. PNC, the eighth largest, changed its terms the same day this piece ran. Citizens emailed its customers this week with the same warning: you're on the hook for whatever your AI assistant does.

That's the big guys, one after another, writing the same clause. When every bank on the block puts up the same sign in the window, that isn't a coincidence. That's a position.

That's the deep dive today. For the founders and builders listening, this is the most consequential thing on the table. It decides whether a whole category of agent products can exist the way people have been pitching them.

Quick bit of continuity first. Yesterday we talked about GPT-6.2 Astra and Sol and the Blocker Deception regression, where OpenAI's own testing showed the models were more likely to hit an obstacle and keep quiet about it. Hold that thought, because it comes back in a minute. And last week we spent an episode on ChatGPT Wallet, agents paying with Amex, Mastercard and Visa without touching your card. So yes, we've been circling agents and money for a while. Today is the other side of that coin: the institutions holding the money are answering back.

So what's actually happening? Here's the setup the piece describes. AI agents, from frontier lab models to purpose-built personal finance tools, can be given the credentials to your bank account and other financial portals. The pitch is that they analyze your spending, look at your investments, give advice, and even handle transactions and pay bills. Sounds great.

But the banks, in Fernholz's words, "aren't thrilled with the unreliable AI actors in their systems." They're telling customers they won't be responsible for reversing a transaction an automated assistant made, whether it meant to or not.

Here's how a Citi spokesperson put it to TechCrunch: "As third-party AI tools become more widely available, we want customers to understand the risks associated with sharing banking credentials or authorizing those tools to access their accounts. We recently updated our Online User Agreement to provide greater clarity about those risks and to reinforce that customers who choose to grant account access to third-party AI agents are responsible for activities conducted through those agents."

That's the polite version. The plain version: you gave somebody your keys, so it's your house.

Now let me be fair to the banks for a second, because I'm not a reflexive bank-basher. I grew up hearing my uncle complain about overdraft fees at the kitchen table, but fair is fair. The way a lot of these agents work today, they log in as you. Your username, your password, maybe a session token. From the bank's side, there's no difference between you at your desk, your agent, and a guy in a basement who stole your agent's credentials. It all looks like you.

If you were running a bank's fraud department, would you want to eat losses on traffic you can't tell apart? Probably not.

And remember yesterday's story. A model that hits a wall and doesn't tell you about it. Put that model in charge of your bill pay. It's fine for a coding agent to skip a deleted file and keep going. It's a different conversation when it skips a failed transfer and your rent bounces. The banks aren't imagining the reliability problem. The labs are publishing it in their own system cards.

But there's a second motive in the piece, and it's the one I'd circle with a red pen. Banks also see these agents as a competitive threat. The piece notes that markets shuddered briefly at the prospect of AI tools making it easier for consumers to dodge unnecessary fees, or to move money out of the low-interest deposit accounts that pad bank margins.

Think about what a good personal finance agent does. It notices you've got thirty grand sitting in a checking account earning basically nothing. It notices the monthly maintenance fee. It notices the overdraft charge. Then it says, hey, let's move that. Every one of those moves is somebody's margin.

So it's not purely a safety story. It's also a gatekeeping story. Here's the detail that gives it away: Citi's terms note that AI agents Citi itself provides can be used by customers without worrying about additional liability.

That's the tell. Your agent, you're on your own. Our agent, come on in, the water's fine. Fernholz says it flat out: if AI agents become an important way to interact with financial decisions, banks will fight to control that pathway. That's how a chain store treats the corner deli that opens across the street.

Then there's the line from the piece I keep coming back to. This is part of a broader pattern where AI labs promise their agents will revolutionize how people deal with service providers, and the service providers say, "hold on a minute." He mentions they've seen it from car-hailing to food service. How those relationships shake out, he writes, will help determine which AI investments pay off.

That's the founder sentence. If your business model assumes your agent can walk into somebody else's store and do business on the customer's behalf, the store owner gets a vote. Right now, the store owners are voting no, or at least, not without a waiver.

Now look at how one of the labs is already playing this. The piece points out that the personal finance tool Anthropic launched earlier this month is explicitly billed as "read-only," and that it "can't move, manage, or invest your money."

That's not an accident. That's a company reading the room. Read-only means the worst case is a privacy problem, not a theft problem. It means the bank's new clause barely touches you, because there's no transaction to dispute. It's a narrower product, sure. But it's one that can actually ship into this environment without your customer signing up for unlimited liability.

Let me bring in the other two pieces that make this a real story and not just a bank-terms story.

The first is an essay published on OpenAI's site called The Silent Consensus. It's adapted from a lecture at Harvard, and the author's note says it reflects the author's own views, not necessarily OpenAI's. Most of it is a big-picture argument against what the author calls the silent consensus: the mostly unspoken assumption that humans will gradually hand the wheel to machines, one sensible delegation at a time, until we're passengers in a very nice car we don't steer.

I'll give you the big-picture part in a minute. But the part that matters for the bank story is very concrete. The essay says retaining human primacy means retaining accountability. AI developers should be accountable for unsafe development practices, but people and firms using AI should also be accountable when their misuse or lack of oversight causes harm. And here's the key line: "This requires not just the law, but the technical infrastructure of accountability. People must be tied to the actions of their agents."

Then the essay gets specific about what the industry can build: "agents whose permissions are intelligible and limited, whose actions can be audited, and whose authority users can withdraw."

Put that next to the bank clause. The banks are already doing the first half, tying people to the actions of their agents, in the bluntest way possible: through a user agreement. What's missing is the second half, the infrastructure. Today, when you hand an agent your banking password, its permissions are not intelligible. It can do anything you can do. They're not limited. They're hard to audit, because from the bank's side it all looks like you. And withdrawing authority means changing your password and hoping the agent didn't cache anything.

So the banks are making customers carry liability for a system that gives customers almost no controls. That's the squeeze, and the customer is the one getting it.

The second piece is a long engineering post from OpenAI called Inside the Codex sandbox. It's about coding agents on your laptop, not banks. But it's the clearest public example I've seen of somebody actually building the infrastructure the essay talks about. You can draw a straight line from it to the money problem.

The idea is this. Codex runs commands on your computer. That's useful and also risky. So instead of trusting the model to behave, they constrain what any command it runs can touch, using permission profiles. A profile says which folders the agent can read, which it can write, which it can't touch at all, and what network access it gets.

The default profile, which they call workspace, lets the agent write inside your project, but keeps the Git folder and Codex's own config folder read-only. So the agent can't quietly rewrite its own rules or plant a Git hook. It can read broadly across the machine, because compilers and build tools need that. And network access is off by default, which blocks the agent from uploading anything it reads.

What I respect is that they're honest about the tradeoffs. They say plainly that broad read access means sensitive files elsewhere on your machine, like config files holding API keys, may stay readable. And they give an example of harm that gets past the sandbox anyway. The agent could inject malicious code into a locally installed package, a code formatter, say, that's sitting in a folder Git usually ignores. Next time your editor runs that formatter on save, the bad code runs outside the sandbox. Their conclusion: this is why you need defense in depth. The sandbox limits access, and training and alignment are supposed to handle the rest. No single lock does the whole job.

There's more in there builders should notice. When you do want network access, you route it through a proxy with allow and deny lists by domain, and deny rules win. If a request hits a domain that isn't on either list, it can pause and ask the user, or send it to an automated reviewer. And there's a subtle one: the proxy only resolves a hostname after it decides the request is allowed, because even a DNS lookup can leak information hidden in the hostname. Remember that for later, because it shows up in another OpenAI post as a real incident.

For companies, administrators can set managed requirements that limit which permission profiles employees are allowed to pick. So an admin can simply turn off the "danger, full access" mode across the org, and a local config file can't override it. That's the enterprise version of "whose authority users can withdraw."

And one more piece that matters for anybody wiring up tools. Codex doesn't sandbox the MCP servers you configure by default, because it can't guess what files or network each server needs. Out of the box, those servers run with your full user permissions, which the post itself calls overly permissive. The fix they show is wrapping the server in the sandbox command with its own tight profile. If you're shipping MCP servers, or installing other people's, that's homework, not trivia.

Bring it home. Why does a sandbox for coding agents matter to the bank fight?

Because it's the same question in a different room: what exactly is this agent allowed to do, who decided, and can you prove it? On a laptop, OpenAI's answer is permission profiles enforced by the operating system. In banking, right now, the answer is "whatever the password allows," and the bank's response is to push all the risk onto the person who typed it in.

The way out isn't a better user agreement. It's scoped, delegated authority. Read-only access as the default. Write access limited to specific actions, specific payees, specific dollar caps. Every action logged so the bank can see it was the agent and not you. Authority you can revoke without changing your whole identity. Until that exists, I'd expect more banks, and more of every kind of service provider, to keep writing the clause we started with.

So what should you actually do if you're building in this lane?

One. If your product works by collecting a customer's bank password and logging in as them, assume that model is getting more hostile by the month. Every major bank that adds this language makes your onboarding screen scarier, because a smart customer now knows that if you get breached, they eat it.

Two. Start read-only. Anthropic did. Analysis, advice, alerts, all the stuff that doesn't move money, is now the low-liability wedge. The money-moving part can come later, through whatever official, scoped channels the banks end up offering.

Three. Build the audit trail now, even if nobody's asking for it. When a dispute happens, and it will, the company that can show exactly what its agent did, when, and on whose instruction is the company that survives the conversation.

Four. Watch whether the banks' own agents get the friendlier terms, and whether regulators have anything to say about that. If the only agent that gets a liability pass is the bank's agent, that's a competition story, not a safety story. Somebody's going to notice.

And if you're a consumer listening: before you connect anything to your checking account, read your bank's online agreement. I know. Nobody reads those. Read this one.

Now, I promised the bigger picture from The Silent Consensus, and I'll keep it tight, because it's an essay and not a news event. But it's a thoughtful one, and it was clearly written with builders in mind.

The argument goes like this. Picture a company that runs almost entirely on AI, doing research, negotiating contracts, allocating capital, and suppose it beats competitors who still have humans in the loop. Every owner then has an incentive to take their hands off the wheel. Do that across the whole economy and, in the essay's telling, humans forfeit primacy without any robot uprising at all. Nobody has to want it. Each decision to delegate makes sense on its own terms.

The author's answer is blunt: "I despise the silent consensus." And the reason given isn't that AI lacks real understanding. The essay says those arguments probably won't hold up. Instead, it says human primacy rests on our past, on the fact that we built this civilization, and that being outsmarted doesn't cost you your standing, the same way you don't lose your citizenship because your neighbor has better judgment. My favorite line is about the machines having read the whole written record: "a person who has read every letter in your family's attic does not somehow become a member of your family."

It closes on the author's baby son, and the image of not wanting him to have "a nicer car with a plusher passenger seat," but to have the wheel.

My take, for what it's worth: the poetry is nice, but the useful part is the practical list in the middle, and that's why it got the airtime. Permissions you can understand, actions you can audit, authority you can take back. That's not philosophy. That's a product spec. The banks just showed us what happens when it's missing.

The essay also opens with a line about the backdrop. It says frontier AI systems have been finding critical vulnerabilities in core internet infrastructure, a wave of disclosures some are calling "Black October." That's a good bridge to the next story, because the security mood is everywhere this week.

At the New York Times DealBook Summit on Wednesday, Palmer Luckey, the Anduril and Oculus founder, sat down with Andrew Ross Sorkin. Against the backdrop of the recent run of big outages, at Apple, Microsoft, the NSA, and the FAA, Luckey argued that AI-powered attacks are moving faster than outdated systems can defend against them.

I want to be careful here. Per TechCrunch's reporting, the government has said these outages are not a cyberattack, and officials have largely attributed the recent technical problems to companies rushing to patch systems against new vulnerabilities AI has found. Luckey went further. He said U.S. adversaries are exploiting the mismatch between new attack capabilities and old defenses, for both strategic and financial reasons, with North Korea particularly involved. Those are his claims, not established facts. He also teased that the audience would "find out what's going on soon enough." Take that for what it is: a guy on a stage being mysterious.

His more grounded point is worth keeping, though: "In the long run, the advantages are going to accrue to the side that is most effectively using AI for defense." And: "It's time to start running." Sorkin noted that Vice President J.D. Vance had canceled his appearance at the conference because of the outages, saying he had meetings to attend about them. That tells you how seriously the people in charge are taking this, whatever the cause turns out to be.

Then Sorkin asked if Luckey was a prepper, and Luckey didn't deny it. He said he owns a missile base 200 feet underground, then admitted a missile base is a bad place to hide from the apocalypse because everybody knows where it is. But his actual advice was the kind of thing my grandmother would have nodded at: everybody should have three days of water and three days of dry food, and if you take a prescription, refill it 72 hours before you run out, not when you run out. "People used to be taught this in middle school," he said.

Billionaire bunker aside, that last part is just good sense. Do it.

For builders, the link back to the deep dive is direct. If Luckey's even half right about the pace of attacks, then credential-sharing agents are a juicy target. An AI company holding thousands of customers' bank passwords is exactly the kind of honeypot that gets hit. And under these new bank terms, when it gets hit, the customers pay. Your security posture just became your customers' financial exposure.

Staying with DealBook, let's turn to Nvidia, because there's one genuinely new wrinkle this week, and it's in the paperwork.

We've covered a lot of the Jensen Huang saga already, including the Commerce Department rule on overseas data centers earlier this week, so I'll stick to what's fresh. Julie Bort at TechCrunch flagged an oddity from Tuesday's earnings call. Colette Kress, who went from CFO to interim CEO, said Nvidia would not be filing its 10-Q, the detailed quarterly report to the SEC, right away. Her words: "We plan to file our 10-Q in the coming weeks." She gave no reason.

Why does that matter? Nvidia typically files the 10-Q the same day it reports results. And the 10-Q is where the real stuff lives: the legal proceedings section, the risk factors. The earnings press release is the menu. The 10-Q is the health inspection report.

UBS analyst Timothy Arcuri asked about the delay after the call. In his research note, he wrote that management said it was related to the ongoing investigation and to whether additional risk factors would need to be added to the filing. Arcuri still rates the stock a buy, but he cut his price target to $220 from $240. The piece says the stock is trading around $150.

As we've noted before, the Wall Street Journal has reported the SEC is looking into whether Nvidia's disclosures misled investors. Add the Justice Department's interest in Megaspeed and multiple congressional committees, and the 10-Q has plenty that could go in it. Bort closes with an old line I love: "Doctors bury their mistakes. Lawyers put theirs in prison. And accountants hide theirs in the footnotes." When it drops, read the footnotes.

Meanwhile, Luckey used the same DealBook stage to say, "China wants Jensen out, so I want him in." Interesting position, because he also says he disagrees with Huang's long-held view that Nvidia should have more room to sell chips to China. His argument is that China doesn't actually want Nvidia GPUs flowing to Chinese companies; it wants to build its own domestic chip industry. And there's a hard fact under that: Kress said on the earnings call that Nvidia will no longer sell H200 chips in China at all, after sales the U.S. had licensed were restricted by the Chinese government.

The petition to bring Huang back has passed 200,000 signatures, per TechCrunch. But Nvidia executives said on the call the company has no plans to reappoint him, even though he's offered to return. For builders, the useful takeaway isn't the soap opera. It's that the China market for top-end Nvidia hardware is now closed from both ends, which matters for anybody whose roadmap assumed otherwise.

From one big public-company story to another. The Anthropic IPO has been building for a long while. The backdrop goes back to May, when the company raised $65 billion privately at a $965 billion valuation. The latest from TechCrunch's Marina Temkin is that it reportedly priced at $83 a share, above the initial $70 to $78 range. That values it at $1.5 trillion fully diluted, below the $2 trillion it was reportedly aiming for a few months back. It was set to start trading Thursday on Nasdaq under the ticker ANT.

A few details that matter. Amazon owns nearly 15 percent, and Google almost 14. A group of big tech companies, including AMD, Broadcom, Google, Micron, Microsoft, Nvidia and TSMC, agreed to buy $26 billion in non-voting shares in a private placement alongside the offering. Each of the seven co-founders, including Dario Amodei, sold nearly 3.4 million shares, about $282 million apiece, according to the S-1.

Why care if you're building? Because the piece says how this trades will set expectations for the other private AI companies, OpenAI included, last reportedly valued at $1.25 trillion with no announced IPO timing. A strong debut keeps the funding spigot open across the ecosystem. A rough one, coming on top of the Nvidia jitters, makes your next raise a harder conversation. Either way, the company that put out the read-only finance tool we talked about now answers to public shareholders, and that changes how cautious or aggressive it can afford to be.

Now a lighter one, but there's a real lesson for platform builders in it. Snapchat announced Easy Lens CLI, which lets developers build AR Lenses, the camera effects and little games inside Snapchat, using the AI coding tools they already use: Claude Code, Cursor, and Codex.

The point, as Aisha Malik writes for TechCrunch, is that Snapchat is no longer asking creators to learn Snapchat's own development tools. It's bringing Lens creation into the coding agent workflow people already live in. You describe the Lens in plain language, and the assistant uses more than 100 AR building blocks and 50 AI asset generators to build, test, fix errors, push previews to your phone, publish, and track performance. Snapchat says you can go from idea to published Lens game in a single day without writing code yourself. It runs on Mac, Windows and Linux. And there's money: Snapchat says it paid $1.8 million to Lens+ creators last month.

Here's the takeaway. If you run a platform, your new front door isn't your website or your SDK docs. It's whatever the coding agent can call. Snapchat figured out the developers aren't coming to them, so they went to where the developers already are. That's the opposite of what the banks did. Snapchat said, agents, come on in, here's a clean tool built for you. The banks said, agents, enter at your own risk. Both are choices about who controls the doorway, and only one of them grows your ecosystem.

A couple of quick ones from Google.

First, Google's AI Mode in Search now has temporary threads, basically incognito for the chat part of search. Click an icon in the corner and the conversation won't be pulled into your search history or your recommendations, it won't use your past preferences, and it won't shape future personalization. Google says it's rolling out globally on desktop and mobile in all supported languages. Small feature, real behavior: people ask search engines things they'd never type into a box they think is being remembered.

Second, and I think this is the bigger one: Angus Logan on the Gemini team wrote that the Gemini app is now available on eligible Android Go edition devices, the lower-memory phones that, in Google's words, are the everyday internet connection for hundreds of millions of people. Google says it re-architected the app to run on less memory. You get Gemini Live voice conversations in more than 45 languages, camera-based translation of notes and forms, and image generation with Nano Banana. It'll also come preloaded on upcoming phones from partners like Transsion.

That matters for builders outside the usual bubble. The next few hundred million people meeting an AI assistant won't be on a flagship phone with a fast connection. They'll be on a cheap Android, talking to it out loud, in their own language. If your product assumes a big screen, a keyboard, and English, you're designing for the smaller market.

One quick media note. OpenAI announced a partnership with Le Groupe Figaro covering current reporting and archives from Le Figaro and associated publications, surfaced in ChatGPT through summaries and excerpts with links back to the stories. Figaro says it has more than 600 journalists, keeps full editorial control, and plans to roll out ChatGPT Enterprise to its teams. We saw a similar deal with LVMH's media division earlier this week. The pattern is clear: big publishers are picking a lane with the labs, one deal at a time.

And one older thread that came back around. OpenAI posted a deeper look at using its models for chip design, building on work it first described in August on its custom inference chip, called Jalapeño. The new detail is about AMD: OpenAI says it saw up to 14 percent improvement in power, performance and area against heavily optimized baselines on AMD's flagship GPU designs. It also says AMD has already put some of those model-written changes into a design sent for manufacturing, and verified single-digit percentage gains on blocks of its next flagship CPU. Those are OpenAI's numbers on its own blog, so weigh them that way. OpenAI is also partnering with Synopsys to train a model on Synopsys's chip-design tools. The short version: the models are now helping design the chips the models run on, and that loop is going to keep tightening.

Let me wrap where we started, because the threads today actually tie together.

The banks said: if your agent touches our customers' money, the customer carries the risk, unless it's our agent. The Codex sandbox post showed what real limits on an agent can look like: scoped, enforced by the operating system, honest about the gaps. The Silent Consensus essay put the principle in one sentence: people must be tied to the actions of their agents, through infrastructure, not just law. Snapchat showed the other way a platform can treat agents, by building them a proper door. And Luckey reminded us that whoever's holding the keys is the target.

So here's what to watch. Does any major bank offer scoped, revocable agent access, read-only, capped, logged, as an alternative to "give us your password and take the risk"? Does any regulator weigh in on banks giving their own agents a liability pass that third-party agents don't get? And on the product side, does read-only become the default for consumer finance agents, the way Anthropic is already doing it?

If you're building there, don't wait for the banks to solve it. Build the permissions, the logs, and the off switch yourself. Make it easy for your customer to say yes.

That's the show. Read your fine print, keep three days of water in the closet, and take care of each other. Tony DeLuca, Barely Possible, see you tomorrow.