September 2026 Summaries

Michael Ferranti argues that financial institutions deploying AI agents need runtime controls, particularly feature-flag-based kill switches, to quickly disable or revert problematic AI-driven behavior without requiring a new code deployment. Citing Gartner forecasts of growing AI security spending and risks from weak access controls and prompt injection, the piece contends that faster AI-generated code releases increase the need for safeguards beyond upstream prompt scanners and AI gateways. It references Google’s and Cloudflare’s outage postmortems as examples supporting broader use of feature flags and global kill switches. In a loan-origination scenario, a kill switch could revert AI approvals to established rules-based underwriting while discrepancies are investigated, and the same approach could govern new capabilities, model changes, and agent write permissions. The piece also recommends assessing existing AI workflows for rapid shutdown capability, authorization processes, auditability, and the resilience and ownership of the kill-switch system, while presenting Unleash’s FeatureOps platform as a tool used by financial-services organizations.