OWASP guidance warns AI agents may keep working after shutdown

An article from The Hacker News, published on 13 November 2026, examines how AI agents that manage workflows can continue working even after a “kill switch” is issued. It notes that an agent is typically part of a broader system, with an orchestrator, credentials and tools that can carry out actions beyond the agent itself. A 2026 OWASP GenAI Security Project roundup is cited, highlighting that containment should be able to revoke credentials and network access, and that outcomes vary by deployment.

The piece stresses that turning off the model does not automatically stop all ongoing work, and containment must consider queued tasks, active sessions and multiple tokens or connections.

The article then sets out a practical response sequence for containment and investigation. Signals warranting attention include unusual access to resources outside the assigned task, abnormally high volumes of reads or exports, workflows continuing after supposed pause, unexpected credentials or integrations, and logs that do not match normal activity.

A five-step response is presented: pause the affected workflow; revoke or limit the agent’s access and rotate any exposed secrets; restrict the involved tools and network paths; preserve evidence and determine what actually changed or accessed; and contain downstream effects while engaging the appropriate owners. The guidance emphasises that evidence should be correlated across identity, API, application and network logs, and that preventing further activity does not undo completed actions. The article also references Keeper Security’s KeeperPAM and Secrets Manager as examples of how to govern credentials behind AI-driven workflows.