Databricks Omnigent adds intent-based authorization for AI agents

Omnigent now supports intent-based authorization, where every action an AI agent takes is checked against a declared purpose. This complements identity-based authorization by ensuring actions are not only permitted but also aligned with the current task, preventing unauthorized actions even if the agent has the credentials.

Intent-based authorization in Omnigent provides three verdicts for tool calls: 'Permitted' for actions within scope, 'Consent-required' for sensitive actions needing human approval, and 'Denied' for actions outside the declared purpose. These are managed through ALLOW, ASK, and DENY constructs.

Users can define agent intents either at design time for autonomous agents or have the agent draft intents from descriptions for interactive agents, requiring human approval before they are applied. This ensures that agents cannot unilaterally change their purpose at runtime.

The new authorization model effectively blocks prompt injection attacks by denying actions that fall outside the agent's declared intent, even if those instructions are embedded within data the agent processes. This provides a layered defense against sophisticated threats.

Intent-based authorization is integrated with other contextual policies, such as session-risk scoring, within a single policy engine. This provides a robust, multi-layered security approach where multiple checks reinforce each other, and any single denial halts the process.