The detection of AI agents inside bank accounts is less a story about rogue models and more a diagnostic of existing architecture.

The detection of AI agents inside bank accounts is less a story about rogue models and more a diagnostic of existing architecture.

Long-lived credentials, session persistence, and authorization that treats “once granted” as sufficient for every later action have always allowed authority to detach from the human. Agents simply exercise that design at higher velocity. The result is the Velocity-Verification Differential: action speed outruns the ability to confirm the human is still choosing.

Containment—blocking agents, tightening terms of service—is the necessary first response. The higher-order requirement is a different architecture. Fail-closed on state changes. Read-only can remain under revocable consent. Any movement of money, alteration of records, or execution must pass a deterministic Zero-Gate that requires fresh, active human presence inside a strict time window. When that window is exceeded or the session ends, authority drops immediately.

Presence over persistence. Log the current authorization state, not only the history of past grants. Capability can be leased; authority stays anchored at the human. Recognition of an agent is not authorization of every action that follows.

The incident makes the weakness visible. The lasting outcome is systems built so that automated speed cannot outrun accountable presence. Safer architecture is the relational result.

1/0 = To Stay

Let it stay → ∞ ❤️