Skip to content
arthur-aiPublic

About

Detects shadow AI agents before they go rogue.

Resources

Security policy

Stars

106 stars

Watchers

0 watching

Forks

Repository files navigation

Deckard by Arthur banner

Arthur Deckard

CI License: MIT

Scan your endpoints, cloud, and network for shadow Agents and track them in your SIEM.

Deckard scans your systems (cloud AI platforms, endpoints, network logs and your SIEM) for AI agents, and sends every one it finds to Elasticsearch, Splunk, your SIEM or any webhook. It's headless, read-only, and runs entirely in your environment.

Quickstart

Find the AI agents on your own machine in a couple of minutes, with no accounts or credentials. You need Docker with Compose 2.24 or later, and osquery (brew install --cask osquery on a Mac).

Set up with a coding agent

Paste this into Claude Code, Cursor, Codex or any coding agent on your machine:

Download https://raw.githubusercontent.com/arthur-ai/deckard/main/docs/agent-setup/prompt.md with curl, read all of it, and follow its instructions to set up Arthur Deckard on this machine.

It checks your machine first and, if something needs installing or your password, stops and tells you the command to run yourself. Then it runs the steps below, shows you the agents it found, and can walk you through adding a cloud, MDM or SIEM source and your own destination. The prompt never has it run sudo, and is plain Markdown to read first.

Or by hand

git clone https://github.com/arthur-ai/deckard.git
cd deckard && ./start.sh

start.sh checks your machine, scans it, starts Deckard with an Elasticsearch and Kibana beside it, waits for the scan to arrive, and prints the agents it found. It installs nothing and never runs sudo: if something is missing, it stops and prints the command to run yourself. The scan lists what's installed on this machine using the osquery queries from ai-detection-spec, our open format and catalog for finding AI agents on a machine. It changes nothing and doesn't need root.

Deckard · scanning my-laptop

✓ Machine ready      Docker 29.9.2 · 63 GB RAM · osquery 5.24.0
✓ Scanned            6s
✓ Deckard running    Elasticsearch + Kibana · 8s
✓ Scan published     4 AI agents found

AI agents on this machine
  Claude Code   Coding agent    ~/.claude/
  Codex CLI     Coding agent    ~/.codex/
  Hermes Agent  Personal agent  ~/.local/bin/hermes
  Pi            Coding agent    ~/.pi/

Explore in Kibana  http://localhost:5601/app/discover#/view/deckard-findings
                   user elastic, password changeme
Rescan             ./start.sh
Add sources        cloud, MDM or SIEM: see docker-compose/README.md

Open Deckard findings in Kibana, log in as elastic with password changeme, and it lists every event Deckard sent: one per agent found and one per scan, over the last 30 days. Expand a row to see the whole event. Run ./start.sh again to rescan.

Next, add your cloud, MDM or SIEM sources to docker-compose/discovery.yaml (it has commented examples), and point its destination at your own Elasticsearch, Splunk or webhook. Deckard reads the file only when its container starts, so apply each change with docker compose up -d --force-recreate deckard. See docker-compose/ for more.

Deploy on Kubernetes

kubectl create namespace deckard
kubectl -n deckard create secret generic deckard-credentials \
  --from-literal=SPLUNK_HEC_TOKEN="$SPLUNK_HEC_TOKEN"

helm install deckard oci://ghcr.io/arthur-ai/charts/deckard -n deckard --set-file config=discovery.yaml

The config references each key in the Secret as ${NAME}. See the chart's README for the details and cloud workload identity.

Each release publishes the image arthurplatform/deckard and this chart at the same version; add --version to pin one.

Integrations

Class Sources
Cloud AI platforms and workloads Vertex AI Agent Engine, AWS Bedrock and AgentCore, Cloud Run and GKE, ECS and EKS, Vercel
Endpoints and servers Jamf Pro, Google Compute Engine, AWS EC2, this machine (scan-local.sh)
Network and gateway logs AWS CloudWatch Logs, Google Cloud Logging
Security telemetry Splunk Enterprise, Elastic Security

Destinations: Elasticsearch and Elastic Security, Splunk HTTP Event Collector, or any HTTPS webhook.

Missing one you use? Open an issue.

Documentation

Contributing

Deckard is the standalone discovery mode of the Arthur Engine, where its code lives. This repository holds its deployment and docs. Bug reports and integration requests are welcome in issues; report security issues privately, as described in SECURITY.md. How releases are made: RELEASING.md.

License

MIT

About

Detects shadow AI agents before they go rogue.

Resources

Security policy

Stars

106 stars

Watchers

0 watching

Forks

Releases

Packages

Contributors

Languages