AI agent intent is a starting point, not a security strategy

This report presents research showing that a large portion of agentic chatbots retain live credentials (65%) and frequently use hard-coded keys (51%), creating orphaned access and credential risks; it also demonstrates how prompt-injection can cascade through multi-agent pipelines, bypassing conventional SOC visibility. The authors recommend treating agent creation as an identity-governance event, adopting delegated authentication, enforcing runtime intent/behavior policies, and improving discovery and enforcement across managed and self-hosted agent frameworks.