Banks Block AI Agents, Plus 2 Tech Stories
Banks are drawing a hard line around AI agents and bank accounts. That sounds like banks being banks until you get to the important part: some institutions reportedly can’t reliably tell whether an automated action was requested by the customer or decided by the software.
Banks block AI agents from customer accounts
Axios reported on October 22 that major U.S. banks have detected unvetted third-party AI agents inside customer accounts. Reuters subsequently reported that Morgan Stanley, JPMorgan Chase and Charles Schwab were among the institutions limiting or blocking that access. The reporting is based partly on unnamed sources, and the banks haven’t said customer money was stolen. That distinction matters. This is a serious control problem, not evidence of a broad bank-account heist.
How does an agent get in? In some cases, the customer gives it a username and password. The software then signs in through the same website or app interface a person would use, rather than through a bank-approved connection with narrowly defined permissions. To the bank, a customer-directed balance check and an agent improvising its way toward a transfer can look uncomfortably similar.
Read-only access is one possible middle ground. An agent could collect balances or categorize transactions without gaining the ability to move money. But even read-only financial data is sensitive, and “read-only” doesn’t answer where the agent stores what it learned, who can retrieve it later, or whether a future prompt can expose it. This is the same permissions problem we saw when AI agents reached live business systems, only now the resource at risk is your money.
I think the banks are right on the narrow point: a general-purpose agent should not receive the same credentials and authority as the account owner. The useful design is boring on purpose. Give software a limited token, let it perform a specific action, require a human confirmation for transfers, and make that authority easy to revoke. Until your bank and the agent provider offer that kind of official integration, don’t hand an AI assistant your banking password. Convenience is not a control.
Turning on AI is the easy part
Deciding what an AI tool should be allowed to see, who should use it, what work it should perform, and what happens when it gets something wrong is the more interesting problem. Raymond Tec helps businesses connect and automate the tools they actually use without treating every new feature like a button that obviously needs to be switched on.
Microsoft’s medical AI asks before it answers
A Microsoft research system is trying to solve a more fundamental problem with medical AI: real patients do not arrive as complete case summaries. Doctors ask questions, examine the patient, order a test, reconsider the possibilities and then decide what information they need next.
The study published in Science calls this sequential diagnosis. Microsoft’s Diagnostic Orchestrator coordinates simulated clinical experts, builds a differential diagnosis and chooses the next question or test based on what it has learned so far. On more than 300 unusually difficult cases, the system outperformed general physicians in the study and designed less costly diagnostic workups. That’s more meaningful than acing a multiple-choice medical exam because the system must decide what evidence to gather, not merely recognize the answer after someone else has assembled the clues.
It is still a research result. The cases were structured benchmarks, the costs were estimates, and a model that handles rare diagnostic puzzles well has not automatically proved that it can triage a frightened person with incomplete records in a crowded emergency department. Microsoft and Beth Israel Deaconess Medical Center plan to test the system for that triage setting next. That hospital work, including what happens when the model is uncertain or wrong, will tell us more than the headline accuracy number.
There’s no consumer action here, and certainly no reason to ask a chatbot to replace a clinician. The promising part is the mechanism: forcing the AI to gather evidence in stages and expose a workup, rather than producing one polished answer from a data dump. In medicine, a confident sentence is cheap. A safe process is the product.
Technology is rarely just about the technology
Some of the most important technology stories aren’t product launches at all. They’re about health, privacy, education, law, accessibility, work, and what happens when technology reaches ordinary people. Browse more Raymond Tec News for the stories worth understanding without the hype.
Greece’s parliament approved a law on October 23 that will bar children under 15 from social-media platforms beginning January 1, 2027. The policy responds to concerns about addictive design, sleep, anxiety and unwanted contact. It also moves responsibility away from a settings page parents may never find and toward the platforms themselves.
The difficult part is enforcement. A platform cannot block under-15 users without estimating or verifying age. That may mean a government-backed credential, an independent age-assurance service, device-level checks, or some combination. Each option can reduce children’s exposure while creating a new privacy problem: proving age without building a fresh database of identities, faces or documents that can leak or be misused.
That tradeoff is the real story. A rule that protects children only by collecting more sensitive data can solve one problem and manufacture another. Greece has until the start of 2027 to show what its system actually requires, how appeals work, and whether children’s information stays out of the advertising machinery. Readers outside Greece don’t need to change anything today, but this is worth watching because age assurance is becoming the enforcement layer for online-safety laws in more countries.
What to watch next
All three stories come down to the same question: what authority does software receive, and who can stop it? Banks are limiting what agents can do with money. Medical researchers are constraining how an AI reaches a diagnosis. Greece is making platforms prove who may enter. The headlines are about AI and social media, but the durable issue is control. If a system can act, diagnose or decide access, its boundaries need to be visible before something goes wrong, not reconstructed afterward from a log nobody thought to read.
Still in a reading mood? The Raymond Tec News archive covers security, AI, small-business technology, policy, and the places technology collides with ordinary life — without requiring a computer-science degree to get through it.
Sources / Further Reading
- Axios: Unvetted AI agents are accessing U.S. bank accounts
- Reuters: U.S. banks are blocking third-party AI agents
- Science: Sequential diagnosis with language models
- Microsoft AI: Building medical superintelligence
- Reuters: Microsoft develops step-by-step diagnostic AI
- Reuters: Greece passes social-media ban for children under 15
- Associated Press: Greece’s planned age limit and enforcement approach
Photo by RDNE Stock project on Pexels.
