Banking Agent Kill Switch How It Works | Moring
A banking agent kill switch works by halting AI activity immediately, so no further agent step takes effect until a named person restarts it. The control supports bank-wide stops or narrower scopes, is held by named staff, operates without vendor assistance, and records who used it and why. Beneath it sit holds, where a single step waits for a signer, and blocks, where a rule-breaking step never runs.
Short answer
A banking agent kill switch works by halting AI activity immediately, so no further agent step takes effect until a named person restarts it. The control supports bank-wide stops or narrower scopes, is held by named staff, operates without vendor assistance, and records who used it and why. Beneath it sit holds, where a single step waits for a signer, and blocks, where a rule-breaking step never runs.
How it works
The mechanism is deliberately simple in concept and demanding in engineering: a control that halts AI activity immediately, with no further AI step taking effect until a named person restarts it. Banking guidance frames this as one of a family of controls, ranked by strength. Holds pause a single AI step until a named signer approves it; blocks prevent a step that breaks the bank's rules from ever running; the kill switch stops all AI work at once or a narrower scope. Together they give a bank real authority over AI in real time, rather than discovering what happened afterward.
Four properties define a usable implementation. The switch must stop all AI work at once or within a defined narrow scope, such as a single agent class, workflow, or business line. It must be held by named staff with defined authority, not an anonymous on-call rotation. It must work without the vendor's help, so a third-party outage or unresponsive provider cannot disable the bank's own containment. And it must leave a record of who used it and why, which converts the emergency action itself into auditable evidence.
Human-in-the-loop design supplies the finer-grained layer beneath it. Action approval gates interrupt an agent immediately before it executes a designated high-risk action and hold execution until a human with the appropriate authority explicitly decides, evaluating the action's resolved parameters and context against policy before any side effect occurs. This addresses the failure class where an agent operates within its granted privileges yet is about to do something materially consequential or hard to reverse. The kill switch and the approval gate are complementary: one governs individual consequential actions, the other provides emergency containment when the pattern of behavior itself warrants a halt.
Runtime governance research explains why banking needs this hierarchy now. Agentic systems initiate payments, submit trading orders, approve credit applications, file regulatory reports, and settle claims at machine speed without per-action human review, and existing control frameworks were designed for human decision-making rather than autonomous execution. When an agent misbehaves at that speed, retroactive review is inadequate; containment must be exercisable in the moment. Even where AI only prepares work, its output flows into downstream processes, so the ability to halt is relevant before any autonomous execution begins.
The audit dimension is where banking practice diverges from generic security. Governance pattern libraries for regulated financial services pair kill-switch capability with hash-chained audit ledgers, risk-state machines, and autonomy ladders, designed to produce the records an examination would ask for. Every halt, restart, and the justification for each becomes an entry in the evidence chain. That discipline matters because an emergency stop that cannot be explained afterward becomes its own finding.
Supervisory mapping follows directly. The EU AI Act's human-oversight expectations for high-risk systems include the ability to intervene and interrupt, which a kill switch operationalizes. SR 11-7 model governance extends to agent-driven processes once halt and restart authority is documented, FFIEC examination themes on incident response and access control apply to who may trigger containment, and DORA resilience programs can rehearse kill-switch activation and recovery as part of operational resilience testing.