Databricks Omnigent Uses Contextual Policies to Block Slow-Burn Attacks

The agent's control over policies is intentionally limited, with no tools to remove or disable existing guards. Adding new policies requires explicit user approval, and combined policies adhere to a 'single denial wins' rule, ensuring robust protection against agent-level manipulation.

Databricks Omnigent now supports stateful contextual policies that monitor session events and maintain a risk score. These policies can dynamically allow, deny, or seek approval for actions based on the cumulative activity within a session, effectively identifying and blocking slow-burn attacks.

Contextual policies in Omnigent can be configured to either outright deny risky actions or to pause the process and request human approval. This flexibility allows for a tailored security approach that balances protection with operational continuity.

The post explains how indirect prompt injection hides instructions within fetched content, and how slow-burn attacks split malicious goals into seemingly harmless, isolated steps. Traditional security measures that evaluate actions individually fail to detect these combined threats, leading to potential data leaks.