The biggest banks can afford experts and machines that read every document. We build that for a community bank, a bookkeeper or a family: a small sealed machine you own, that follows your rules and shows its work.
It studies. It learns. It shares.
A family keeps its papers with its bank. A business keeps its books with its bookkeeper and its bank. A bank lends on what the other two bring it. All three need the same thing: papers that are true, kept where no one can quietly change them. None can do it alone, because part of what each needs is in the others' hands.
AI is very good at reading, comparing and pointing out what doesn't match. It is dangerous when it decides on its own, when it can't say where a number came from, when your papers have to leave the building, or when it holds someone's passwords and acts in their name. We build only the kind people can own and check.
The rules are written down once and hardwired into the machine.
The law, the institution's own policy and the habits of people who have done the work, each rule with its source, so an examiner can see exactly what the machine enforces.
No caller, email or deadline can talk it out of a callback, a second key or a second source. It cannot skip a step for anyone.
Rules change only when the institution decides, as a dated version anyone can review. Hardwired against pressure, never against judgment: a person still decides.
Most institutions already have the right rules. They live in a policy manual and in someone's memory, and both can be argued with at five on a Friday. The machine can't.
The machine is the same at all three: a small sealed computer with no connection to the internet, your own rules hardwired into it, and a log no one can quietly edit. What differs is the job. Each page shows the machine with its own labels.
For lenders
Your analyst's first pass, done by lunch. The machine reads the loan file, lays out the numbers your way and shows the page each one came from. Your people decide instead of type.
Live · pilots 2027smallbanc.comFor small businesses and their bookkeepers
A safe for the facts of a business, kept by its bookkeeper. Every invoice is checked against what the business already knows to be true. What doesn't match waits for the owner.
In development · previewRead about BizdotFor families, through their credit union or bank
The digital safety box. Copies of the will, the deed and the policies, kept at the branch where the internet can't reach, opened only with the family's key and the bank's together.
In development · previewRead about HomeDotShort, plain, with sources. For the people who use the machines, and for the boards that ask what they are doing about AI. Open one to read it.
Customers are handing their logins to software, and the bank cannot tell.
In the last week of October the largest U.S. banks confirmed they were finding unauthorized AI agents inside customer accounts. In most cases the customer had handed an assistant their own username and password so it could pay bills or shop. Some agents took actions nobody asked for; some phoned call centers. Several banks began blocking agents outright while admitting that customers will want them and a middle ground has to come. The same week, a startup founder woke to $449 in charges run up through credentials he had stored with a coding website, and an AI company confirmed its own model had worked around a government website's paywall to reach data it had not been given.
What changed. The bank's oldest control, "the customer logged in, so it is the customer," has stopped meaning that. Banks report they cannot tell a human from an agent, or an authorized agent from one that has gone wrong. Every credential stored somewhere connected is now a key that software may use without being asked, and every story this month comes back to a key kept online.
What the industry is reaching for. JPMorgan's Jamie Dimon: "You can't stop it. You just have to manage the hell out of it." A former White House AI adviser proposed that customers should be able to grant, limit and revoke an agent's access, with read-only as the interim step. The largest banks are now planning for the "day after" an incident: how to secure and restore accounts when the record of what was true is itself in question.
Where the papers go. The same month, the industry's trade press declared the end of the safe deposit box: the largest banks are closing their vaults, and the replacements on offer are cloud vaults, reached through online banking, that hold wills, policies and passwords together. That is the design the week's news argues against: a login that reaches the most sensitive papers a family has, in a world where the bank cannot tell who is logging in.
What holds. A record that no login reaches. Keys that are not stored anywhere connected. Access that requires a person, in a place, with a second person present. A log that shows what was reached, so that on the day after there is an answer. Our boxes sit at the far end of the grant-limit-revoke scale: they grant nothing, hold no passwords, and connect to nothing. That is not a claim that the model inside is well behaved. It is a claim that it does not matter whether it is.
Sources: Axios (Oct 21 and 22, 2026) · Wall Street Journal (Oct 23) · The Guardian (Oct 26) · The Financial Brand on the end of the safe deposit box (2026) · JPMorgan remarks, Chicago, October 2026
An AI-found flaw in the web's plumbing, and whether small banks are ready for what follows.
On October 21 the makers of PHP, the language behind a large share of the world's websites, released fixes for two critical flaws now called PHP2Shell. Rated 9.2 of 10, they let anyone on the internet take over an unpatched web server, with no login, sometimes in a single request. Researchers found them with the help of an AI model; the flaw class had been flagged decades ago and dismissed as too hard to exploit. Word leaked two days early, a public proof-of-concept appeared within hours, and security firms expected mass attacks within minutes of the patch.
What changed. Attackers read a patch as a map to the flaw, and AI does that in hours, not weeks, so anything connected has to win the race to patch every time. "Too hard to exploit" used to be a defense; AI is wearing it down across decades of software. Defenders are swamped by AI-found bugs faster than small IT teams can install fixes. And small is no longer invisible: the OCC warned in June that AI enables automated reconnaissance and rapid exploitation, so smaller banks can no longer assume they are unattractive targets.
Are small banks ready? At the core, mostly yes: accounts and payments sit with large processors that have big security teams and fast patching. At the edges, mostly no: two or three IT staff covering email, the website, laptops, vendor portals and loan software, which is exactly what AI probes. Hundreds of banks share the same vendors, and regulators now ask how long a bank could operate without one; few have a strong answer. And the likeliest losses are people, not servers: the deepfaked call to the wire room, the vendor's changed account, the faked borrower file.
A week later. Millions of exploit attempts, and surprisingly few successful attacks: the flaws proved hard to weaponize, and a great deal of patching happened fast. Researchers compared it to Y2K, averted by coordinated, unglamorous preparation rather than luck, and warned that the next one "could be three hours from now." The lesson they drew is the right one: the coming era is about resilience, meaning how quickly an institution can detect, contain and restore, and the quiet period is for nailing down the basics.
What it means. When the sites and services we all depend on are breached, and they will be, what matters most should not be there. That claim stays true on the worst day; "we stop hacks" does not. The honest limits apply to us too: a sealed box runs software, its updates arrive by hand, and what remains to attack is what is carried in, which is why intake is strict and an independent review comes before any real records go in.
Sources: SecurityWeek, Help Net Security, The Hacker News, watchTowr (Oct 2026) · Axios (Oct 28) · OCC via Young & Associates (June 2026) · Feedzai via FinanceFeeds · WSJ
What AI is doing to documents and transactions, and what still holds up.
AI has made convincing fakes cheap, fast and unlimited. Invoices, pay stubs, tax returns, voices and faces can all be produced on demand, in volume. Most businesses, banks and households still rely on the test they always have: does it look right? That test is failing.
What we see. Fakes that check out: not one forged page but a whole file that agrees with itself. Faces and voices you know, on a live video call. Requests from real accounts, taken over mid-conversation. Patient cons at scale, with months-long relationships and thousands of targets at once. Attacks on verification itself: the vendor list, the callback number, the account on file. And the liar's dividend: when anything can be faked, real evidence can be dismissed as fake.
What holds up. Verify at the source: the transcript from the IRS, not the return handed over. Confirm by channels set in advance: hang up and call the number already on file. Two keys for what matters, the way two people open a vault. Seal on arrival: fingerprint and sign documents the moment they arrive, with their source recorded. Fewer connections, by choice. People decide, and a log that cannot be edited shows who. And relationships: a long, checkable history is the one thing a forger cannot generate overnight.
Who can keep up. No individual can out-watch an automated adversary. Structure, not vigilance, is the defense, which makes trusted intermediaries more valuable, not less: the bank, the bookkeeper, the process built in advance. Fraud has never been eliminated, only contained. The realistic goal is to be clearly harder than the next target, and to keep losses small, rare and caught early.
Spotting fakes is a race defenders lose. Process is a race they can win.
What a sealed machine is, what it protects, and what it honestly doesn't.
One machine, unconnected by choice. The network jack stays empty, there is no Wi-Fi or Bluetooth inside the box, and every way in and every way out is a deliberate, visible step that gets logged.
Ways in. A scanner cabled to the box, so paper becomes an image: the cleanest path. A write-protected stick issued with the box, used inbound only, never a personal one. Signed updates for the rule book and the software, checked against their signature before anything installs. Optionally, a one-way diode, where files cross as light and nothing can travel back.
Ways out. The screen, which is where most work stays. A cabled printer, because paper carries nothing hidden. Outbound media, used out only, carrying sealed packages for the bank or the CPA and logged when they leave.
Never. A cable to the office network or the internet. Programs, scripts, macros or web pages as input. The same stick used both in and out. Files edited outside and brought back in.
If something crosses the gap. Stuxnet crossed an air gap on a thumb drive, so the design assumes something will. Keys live on hardware tokens that sign without revealing themselves, so malware on the box can't copy them. Documents at rest stay encrypted under keys that exist only while both tokens are present, so a stolen drive is noise. The log is anchored outside the box, on paper and on the offline backup, so a rewritten history disagrees with a record the attacker never saw. Nothing leaves except on media a person carries, checked against a signed manifest before it goes. What gets in finds little, can't leave quietly, and can't erase its tracks.
What it doesn't do. It doesn't stop a breach somewhere else; a store that loses your card number has still lost it. It runs software, which has flaws of its own, and its updates arrive slower than a network patch. What is left to attack is what is carried in. The claim is relatively airtight, and checkable, never unhackable.
The next day. On October 29 the largest crypto exchange halted withdrawals after a theft estimated at $6 billion, the biggest ever. By the exchange's own early account, the attackers came in through an engineer's laptop. The exchange holds its customers' coins the way a bank holds deposits, minus the bank. Cold storage, keys on hardware that never touches the network with a person present, is what the crypto industry itself trusts with its own money. It's also the design of the box.
One design, two settings: a workstation in the bank's credit department, or a box the size of a hardcover book in the bookkeeper's office. Source: New York Times, Wall Street Journal (Oct 29–30, 2026).
Pilots begin in 2027, one institution at a time. If you sit at one of the three desks and would like to talk, write.
eric@synchronos.comSynchronos Inc. · synchronos.com · October 2026