Redpanda introduces the Out-of-Band Policy Engine (OBPE), a governance layer for AI agents that operates outside the agent's reach — meaning agents cannot read, modify, or circumvent the policies enforcing them. The core argument is that in-process guardrails (system prompts, guard models, in-framework permission checks) are fundamentally insecure because a compromised or malicious agent can alter them. OBPE enforces field-level data masking, attribute-based access control (ABAC), and tool argument clamping at the MCP server boundary, server-side, on every call. The release also includes an Agent Network View (live graph of all agents, models, MCP servers, and tool calls derived from traffic), bring-your-own-agent transcripts, AWS Bedrock Guardrails integration, scheduled agent runs, and a CLI. A kill switch to cut off misbehaving agents is announced as coming soon. The entire system deploys into the customer's own cloud account with no vendor-hosted copy of transcript data.

13m read timeFrom redpanda.com
Post cover image
Table of contents
Fixing the agent doesn't work #See every agent in your environment #Govern all those agents #Killswitch: stop bad things from happening #How to tell actual governance from security theater #Early preview features #Nobody built HR for agents. So we did. #
276 Impressions