Agent Blackbox

Agent Blackbox, also branded Guardian Agent Blackbox, is an open-source security plugin from Umanitek for AI agents that can run commands, open files and install packages. It sits between the agent and each action, names what the agent is about to do, checks it against threat intelligence and flags or blocks it. It protects Hermes agents out of the box and OpenClaw workspaces through a bundled bridge.

What It Is

Agent Blackbox is a runtime guard for autonomous agents. Five Hermes hooks observe prompts, tool calls, commands, files, packages and skills before the agent acts. Each action is turned into a deterministic identifier (for example dep:npm:evil@1.0.0) and looked up in a locally compiled ruleset. Categories it covers include prompt injection, dangerous commands, sensitive-file access, secret exposure, malicious or vulnerable dependencies, suspicious skills and known-bad indicators.

How Detection and Enforcement Work

The default mode is audit, which records findings without stopping anything. In block mode, verified threats at or above a configured severity are stopped, along with the user's protected paths and secret exfiltration. Community and heuristic findings only flag. Lookups run locally on the hot path, and the README states that every boundary fails open so a Blackbox fault does not break the agent. Findings appear in a loopback-only local dashboard and an audit trail.

Verified and Community Threat Graphs

Blackbox draws on Umanitek's curated verified graph and an open community graph, both on the OriginTrail Decentralized Knowledge Graph and read through a local DKG node. Sharing to the community graph is opt-in. Reports are signed and use a closed schema with no field for prompts, commands, paths or file contents, and secrets are never shared. Commands let users retract reports, dispute community threats, export their data or erase their reporting identity.

Setup Path

Installation is a one-line script for macOS/Linux, with a PowerShell installer for Windows. Docker is required for the default Blazegraph store, or users can pass --store oxigraph. The installer sets up Hermes, the Blackbox plugin, a local DKG node, the graph subscriptions, attaches protection to local agents and starts the dashboard. Manual install steps are documented, and the plugin requires Python 3.11-3.13.