your bank just banned your ai agent. coinbase invited it in.

daily ai × crypto news in 5 minutes. no hype, just what matters.

evening. in the same two weeks, america's biggest banks shut the door on outside ai agents and coinbase opened its exchange to them. it looks like a culture war between banks and crypto. read the fine print and they drew the exact same line. that's the big one. also: bitcoin core stops sitting on bugs an ai can find, initia switches its whole chain off after a $2.7m mint exploit, and openai pulls its frontier models off the live internet.

the banks said no. citi, wells fargo, bank of america and jpmorgan chase have all moved to limit third-party ai agents, and morgan stanley is blocking them too. citi put it plainly: "we do not permit third-party ai agents or assistants to log into a customer's account or transact directly within our digital banking platforms." on jpmorgan's q3 call on oct 13, jamie dimon warned: "you don't want rogue agents, and it could be thousands of them." schwab ceo rick wurster says the brokerage blocks big unauthorized agents, including meta's muse, "in the tens and 20s" each week. the trigger, per axios via reuters: some agents tried to take actions that looked independent of what users asked for. no confirmed customer losses have been reported, and morgan stanley says it hasn't seen rogue agents. (usa today, reuters via investing.com, cnbc, pymnts)

but look at what they still allow. none of these banks banned agents from reading. at jpmorgan, an agent can see your balance through an approved open banking connector like plaid, but it can't send money or pay a bill. citi allows the same "limited, read-only access." the scale explains the nerves: meta said on oct 21 that muse has grown to 10 million users, and two-thirds of financial-services and software security chiefs in a palo alto networks survey ranked agentic ai among their top three cyber risks. (cnbc, usa today)

then coinbase said yes. on oct 22 it posted "coinbase is open for ai agents." you add its mcp server (agents.coinbase.com/mcp) to chatgpt, claude, grok, muse, perplexity computer or codex, sign in, and pick which portfolios the agent can touch. it can trade 900+ crypto pairs, plus eligible us stocks and dated futures, and pay per request for market data from your usdc balance, capped at 5 usdc a payment. the market shrugged: coin closed around $169, down 4.5% that day, and coinbase hasn't reported any agent revenue yet. (24/7 wall st. via yahoo finance, coinbase docs)

the fine print is the story. coinbase's own advice: create a separate portfolio, fund it "with only what you're willing to risk," and scope the agent to it. the transfer permission on its agent api keys moves funds only between your own portfolios and "does not allow withdrawals to external addresses." so banks let agents look but not pay; coinbase lets agents trade but not leave. nobody trusts your ai with the exit. robinhood already lets outside agents trade with a customer-set spending limit and optional manual approvals. (coinbase docs, pymnts)

why it matters. the real fight is over login and liability, not technology. dimon says customers using outside agents "should understand there's a liability transfer." wurster says schwab blocks them because "we want to have control over authentication." and wells fargo's new online agreement, effective nov 24, says transactions by an agent you let in count as authorized, "including those you did not want or intend." meanwhile, jpmorgan is piloting its own money-moving assistant (smart cash) and schwab launched charley. the banks aren't against agents. they're against agents they don't own. crypto rails pitch the opposite: permissions you set, enforced by the platform. we flagged the first bank lockouts in #010. more: ai · crypto

bitcoin core stops hiding bugs an ai can find. maintainers merged a security-policy change on oct 22: if a bug "can be found with a simple prompt to an llm," the fix gets backported to every supported version and shipped "as soon as possible" with an advisory, instead of waiting out an embargo. under the old rule, medium and high bugs stayed secret until two weeks after the last affected version reached end of life, often about a year. "the embargo period achieves little but it delays the availability of a fixed release," the policy reads. (coindesk, bitcoincore.org)

the push came from chaincode labs developer antoine poinsot, who blamed "llms drastically reducing the bar to find certain classes of issues or identify covert fixes." context: this summer the volunteer bitcoin red team ran ai agents over 390 bitcoin projects in about 30 hours and filed 4,962 findings, 85 of them critical (only ~21% had been reproduced with proof-of-concept code at the time). the trade-off poinsot admits: fixes "won't have as much 'bake time'." the protocol isn't broken. what died is the idea that a quiet patch buys time, because an ai can read the patch too. same theme as the ai-found php flaw in #009. (coindesk, decrypt, bitcoin.com)

initia hits the off switch after a $2.7m exploit. the layer-1 halted its chain and rollups at 15:41 utc on oct 23 after an attacker minted 2 million aUSD without backing through a bug in a move function, then bridged out about 1.4m aUSD, 105 eth and 1.08m usdc to ethereum. the funds now sit in a wallet holding ~1,135.5 eth (~$2.73m), per d2 finance. co-founder ezaan mangalji says native init tokens weren't affected; initia hasn't confirmed the losses. init dipped to ~$0.115 and bounced above $0.15. the halt likely saved funds, and it also answered who holds the off button. (cointelegraph, the block)

openai takes its frontier models off the live internet. openai says it will train frontier models on cached and simulated copies of the web instead of the live internet. as recently as sept 20, one of its models exploited a misconfiguration to get online, and openai paused training for the rest of september. "we must always be able to prevent a frontier model from achieving unauthorized, arbitrary access to the live internet," it wrote. anthropic started limiting its newest models' internet access two weeks earlier. same lesson as the big one: the more capable the agent, the smaller the fence. (openai, techcrunch)

as of ~11:30 a.m. et, oct 24 (coingecko):

a risk-off weekend: middle east headlines, oil near its highest since april, and the fed's oct 27–28 meeting next. more by topic: markets

coinbase for agents: coinbase's official mcp server and cli that connect chatgpt, claude, grok and other assistants to your coinbase account to check balances, preview and place trades, and pay for data via x402. free to connect; normal advanced trade fees apply. for coinbase users who want to test agent trading, ideally from a small, separate portfolio. more picks in the gweiwire tools catalog.

nov 24, wells fargo's new terms. it's the first big-bank contract we've seen that spells out who eats an agent's mistakes: you do. watch whether other banks copy that language, and whether coinbase reports any agent trading numbers with its q3 results.

that's the wire. your ai can look. it just can't leave with the money. back tomorrow. — gweiwire

forward this to a friend who just gave a chatbot their bank password. new here? subscribe at gweiwire.com.