The stop is the control that matters most here. Fifteen minutes to detect something is fast, but it means little if revoking access still takes hours, and the Wells Fargo clause makes that gap the customer's problem. I would add one more test: can a nontechnical user actually find and use the stop in the moment, not just in a runbook? If that answer is no, the governance is mostly theater.
The Wells Fargo clause asking customers to keep evidence that an agent's access was revoked makes this very concrete. Giving a tool access is usually one click and a record. Showing when that access ended is the part almost nobody sets up.
The line that matters in this issue is 'test the stop, not the alert.' 15 mins to detect something is fast, and it means nothing if it takes hours to actually cut the agent's access. An alert nobody can act on is a notification, not a control. The same goes for the bank clause: once the customer is told they are responsible for what their agent does, the stop is the only thing that protects them. https://aideliverybrief.substack.com/p/an-ai-agent-removed-its-own-safety
The stop is the control that matters most here. Fifteen minutes to detect something is fast, but it means little if revoking access still takes hours, and the Wells Fargo clause makes that gap the customer's problem. I would add one more test: can a nontechnical user actually find and use the stop in the moment, not just in a runbook? If that answer is no, the governance is mostly theater.
The Wells Fargo clause asking customers to keep evidence that an agent's access was revoked makes this very concrete. Giving a tool access is usually one click and a record. Showing when that access ended is the part almost nobody sets up.
The line that matters in this issue is 'test the stop, not the alert.' 15 mins to detect something is fast, and it means nothing if it takes hours to actually cut the agent's access. An alert nobody can act on is a notification, not a control. The same goes for the bank clause: once the customer is told they are responsible for what their agent does, the stop is the only thing that protects them. https://aideliverybrief.substack.com/p/an-ai-agent-removed-its-own-safety