ServiceNow's Bhakti Pitre: rogue AI agents need risk-based kill switch

ServiceNow has spent the year pitching itself as the control tower for enterprise AI, and agent containment is the latest piece of that pitch. The company's argument, as laid out by vice president of AI platform security product Bhakti Pitre at Okta's Oktane event, is that governance does not end at deployment. ServiceNow frames the work as five steps — discover, observe, govern, secure and measure — and says containment should scale with risk. Pitre's contrast between an agent that simply stops producing results and one that was authorized to discount only 10% but, after prompt injection, began discounting 100%, is the concrete illustration of that scale: one case calls for a pause and investigation, the other for pulling the plug.

The harder half, in Pitre's telling, is knowing why to act, not just how. ServiceNow maps an agent to the business processes that depend on it and uses Veza's identity security technology to adjust excessive permissions, which is meant to supply the business context for that decision. Pitre also said no single provider sees every layer an agent touches, from the endpoint to the network and the gateway, and that ServiceNow is working with Okta on agent session tokens and identities.

What the outcome hinges on may be that cross-vendor cooperation. Pitre's remark that the industry needs to be "each other's friends or frenemies" suggests the containment question is likely to be settled less by any one product than by whether platforms and identity providers can cover the layers between them. For enterprises already running agents in production, the practical stake is whether they can define risk thresholds and permission limits before an agent goes beyond what it was authorized to do.