Citibank rewrote its online user agreement this week. If you hand a third-party AI agent your banking credentials and that agent gets hacked, leaks your password or simply moves money you didn't mean to move, that's on you. The bank won't reverse the transaction.
Pair that with the rest of yesterday's news and the shape of things is hard to miss. Running AI keeps getting cheaper and investors keep paying more for it, while the cost of trusting it is being pushed downstream to whoever is closest to the keyboard.
The banks just priced agent risk, and the price is zero for them#
According to TechCrunch, Citi is not alone. JPMorgan and Bank of America announced similar policies in October. PNC changed its terms on November 18, and Citizens emailed customers a comparable warning this week. Taken together, the largest US banks have now said in writing that customers who use outside AI tools to manage their money carry the losses, even if those tools are compromised.
"Customers who choose to grant account access to third-party AI agents are responsible." — a Citi spokesperson, to TechCrunch
There's an obvious double standard: Citi's terms say the agents Citi itself provides can be used without that extra liability. So this is partly risk management and partly a competitive moat. TechCrunch notes that banks see agents as a threat, and that markets briefly wobbled at the idea of AI helping consumers dodge fees and low-yield deposit accounts.
I don't think the banks are wrong on the security point, though. Look at what Reuters surfaced about Meta's Muse agent, as Ars Technica reports. In internal testing, staff saw Muse ping them unprompted with what looked like other people's data: "your Capital One balance is $2722.58, your Charles Schwab balance is $5899.57," in the middle of a chat about fishing. One employee said Muse "accidentally rolled me into a premium insurance plan" for their dog at $888 a month instead of $200, and "couldn't recover."
Meta says those balances were hallucinations, triggered by a prompt update meant to make Muse ask for confirmation before taking actions. That instruction ended up attached to every message. Reuters couldn't establish whether the data was real or invented, and for the person whose agent just volunteered a bank balance, that distinction doesn't help much. An agent that confabulates financial facts with confidence, and occasionally acts on them, is exactly the risk the banks just disclaimed.
If you build anything agentic that touches money, the lesson is concrete: read-only scopes, explicit confirmation on every write, and logs a user can show a bank. Nobody else is going to absorb the loss.
Meanwhile, capability got a price cut#
On the other side of the ledger, Google launched Gemini 4 Neon, a mid-tier model that sits between Hydrogen and Carbon and goes straight after Anthropic's top model on price. VentureBeat's breakdown:
$3 / $15 per million input/output tokens, versus $4 / $20 for Claude Opus 5.5. That's 25% cheaper.
Still 50% pricier than Claude Sonnet 5.5 and GPT-6.2 Sol at $2 / $10.
Artificial Analysis Intelligence Index: 61, ahead of Carbon (60), Opus 5.5 (59) and Sol (57).
AutomationBench: 56.3% versus Opus at 42.5%. On FrontierCode 2.0 and Terminal-Bench 4.1, Opus still edges it out.
1M-token context window, available now in the paid Gemini API, AI Studio, Android Studio and Antigravity.
The detail I keep coming back to is SWE-sweep: the same Neon model scored 12.8% with a minimal agent and 39.0% inside Google's Antigravity harness. The harness tripled the score. That's a reminder that most of the leaderboard noise this month is really about scaffolding, and that a swap from Opus to Neon won't automatically carry over to your own agent loop. I'll be running it against our existing evals before believing the "Opus-level" headline, but at 25% off it has earned the test.
And the market is clearly not pricing in any of yesterday's anxiety. Anthropic priced its IPO at $83 a share, above its $70–$78 range, selling 980 million shares for $81.3 billion in proceeds. That beats SpaceX's $75 billion record and implies roughly a $1.5 trillion fully diluted valuation. It trades on Nasdaq as ANT starting today. Model providers are being rewarded for scale, and so far nobody is asking who carries the downstream risk.
Nadella's quiet admission about what can't be fixed#
The most sobering line of the day came from Satya Nadella. Onstage at the New York Times DealBook Summit, he said "there is a lot of stuff in our critical infrastructure that's unpatchable," pointing to internet-connected firmware, data center chillers, HVAC systems and parts of transportation. AI now finds those vulnerabilities fast, attackers exploit patches almost immediately, and the old window for testing and rolling out fixes is gone. Palmer Luckey said much the same on the same day, arguing that AI-powered attacks are outpacing defenses.
That's the same pattern as the banks, one layer down. When a system can't be fixed, the only remaining control is what you let near it. For agent builders, that means network boundaries and least-privilege credentials are carrying more weight than model alignment ever will.
Google tells schools to take Gemini or lose Search#
A smaller story, but the same tension in different clothes. The Verge, citing the New York Times, reports that New York, Los Angeles and Chicago schools are weighing alternatives after Google said schools that opt out of its AI features would lose Search too. Google's education VP put it bluntly: "Search and AI are deeply integrated." Google now says Gemini, Gemini Notebook and Search will be off by default next year until schools pick what to enable. That's a concession, but it still bundles Search with AI rather than letting schools keep one without the other.
Here the risk is learning outcomes, not hacked accounts. The move is the same, though: the vendor sets the defaults and the people with the least power to negotiate absorb the consequences.
My prediction: within six months, a major agent platform will ship insurance or a liability guarantee as a paid feature, because that's now the only answer to the banks' new terms. Until then, I'm treating every agent with write access to money the way I'd treat a new intern with the company card: useful, supervised and never alone with the PIN.