US bank regulators reportedly ask whether cyber-defense AI agents have an emergency shutdown

Take a bank security team that lets an agent close out low-grade alerts overnight. The reported questions apply to that night's work. The team would need to say what the agent did, whether a person approved it, how far its authority reached and how anyone could have stopped it ^[2].

The sourcing is thin. Gizmodo is relaying a Reuters story built on anonymous people ^[1], and its account describes questions and no rule, deadline or penalty ^[2]. A team that builds to these questions is guessing at the form the evidence should take.

The shutdown question has the least obvious answer. Andrew Bailey, the Bank of England governor and chair of the Financial Stability Board, wrote in a letter to the G20: "The global financial system is highly interconnected, and cyber disruption can spread across jurisdictions through common technology providers, shared infrastructure, and cross-border financial activity." ^[5] If an agent runs on a provider's platform, the emergency shutdown may sit on the provider's side of the contract. In our view that is the first thing to confirm for any hosted agent.

We would sort each agent on two axes. The first is whether it only reads (parses logs, scores alerts) or changes things (isolates a host, blocks an account). The second is whether the team can produce a record for each reported question: a log of actions, a named approver with a timestamp, a written limit on authority, and a shutdown someone has tested. A read-only agent with records is done. A read-only agent without them can wait. An agent that changes things and has records needs a recurring check that the records stay current. An agent that changes things without records is the one to restrict first. The cost is speed: every response that waits for a named approver takes longer, and the timestamp on that approval is the evidence.