A payment can look approved and still be wrong. A banking service can come back online while showing stale information. A digital signature can be valid even when the instruction it carries is false.

  1. Who or what can act?
  2. Which record governs?
  3. How do we know recovery is clean?

The questions apply to people, ordinary software and AI agents. Connected software adds speed: it can assemble an instruction, use a credential, call another system and continue through several steps while a control watches only one screen or request.

Who can turn an instruction into a result?

A proposal, an approval, an authorization and an execution are separate events.

Imagine a payment. One system prepares it. A person or automated check reviews what appears on screen. A separate control should decide whether this actor has authority over the actual account, beneficiary, amount and time. An execution service then sends the money.

Finance already uses these separations. A staff member may prepare a payment without being able to release it. A trading system may build an order without permission to submit it. A reconciliation service may find an error without being allowed to rewrite the ledger.

NIST says software agents should have their own identities, credentials and entitlements.[1] This makes delegated power visible and removable. An agent working for a developer, trader or service account should receive only the authority needed for its task.

Bitget reported approximately $387.5 million in attacker-directed transfers in September 2026. Reporting on chief executive Gracy Chen’s account said spoofed transaction information entered the authorization-signing process, while private-key compromise had been ruled out.[2][3]

A valid signature shows that a key signed a message. It cannot establish on its own that the destination, amount, asset and underlying business instruction are genuine. The signing mechanism may work correctly while the institution sends the wrong transaction.

Which record wins when screens disagree?

One completed action can leave records in many places: the initiating application, approval system, bank, custodian, execution venue, transfer agent, depository or settlement service. During an outage, those records may arrive late or show different states.

An authoritative record is the record accepted as controlling for a particular purpose under the relevant legal, contractual and market rules. The first dashboard to reappear does not automatically become the controlling record.

A shared-core banking outage in September 2026 showed why this matters. Affected credit unions reported unavailable online banking, last-known balances used for some decisions, held electronic transactions and step-by-step restoration across more than 300 institutions. The institutions attributed the outage to equipment and cooling failure.[4][5]

During such an outage, the practical questions are concrete: Which record controls the available cash? Which one proves ownership? Has an instruction been accepted, held or settled? What separate evidence can reveal an error in the controlling record?

How do you prove recovery?

Detection, interruption, reversal and final settlement happen at different times. An alert may arrive after work has entered a queue. Ending one process may leave transmitted actions moving through another system. Correcting a customer’s balance may leave liquidity, tax, disclosure, allocation or third-party consequences to resolve.

Global rules for payment and settlement systems require a clear point at which a transaction is final.[6] That makes the timeline important: when activity began, when a machine detected it, when a responsible person learned about it, when authority was withdrawn and when the destination confirmed that it had stopped.

Restoring the service is only the start. The institution must show that its identities, permissions, transactions and records are complete, in the right order, free of accidental duplicates and checked against separate records it can trust.

FFIEC guidance covers integrity across production data, backups and replicas, including the risk that corrupted data can spread into recovery copies.[7] Restarting the same environment cannot prove which state deserves trust.

Before ordinary processing resumes, someone should be able to name the actor, its permitted scope, the route it used, the independent control that could refuse it, the record that governs, the person who can stop it and the evidence that all affected records now agree.

Sources

  1. NIST, identity for agentic AI: https://www.nist.gov/blogs/cybersecurity-insights/back-future-why-agentic-ai-needs-strong-identity-foundation
  2. Bitget support notice: https://www.bitget.com/support/articles/12560603896108
  3. CoinDesk, account of the Bitget incident: https://www.coindesk.com/markets/2026/09/25/bitget-s-usd351-million-hack-happened-via-spoofed-transfers-not-private-keys-ceo-gray-chen-says
  4. WCCFCU, banking-system outage: https://www.wccfcu.com/banking-system-outage
  5. Badger Globe Credit Union, banking-systems update: https://www.badgerglobecu.org/important-update-about-our-banking-systems
  6. BIS, CPMI-IOSCO principles summary: https://www.bis.org/fsi/fsisummaries/pfmi.pdf
  7. FFIEC, data backup and replication: https://ithandbook.ffiec.gov/it-booklets/business-continuity-management/iv-business-continuity-strategies/iva-resilience/iva3-data-backup-and-replication

Methodology · Corrections · Archived original