Guardian Agent Blackbox Is Live: Open‑Source Runtime Monitoring for AI Agents
AI agents can run commands, access files, and install software. That power makes them useful and makes them targets. Blackbox watches every action an agent takes and flags the ones that put you at risk, in real time.
Guardian Agent Blackbox is now live, installable with a single command: curl -fsSL blackbox.umanitek.ai | bash. It's open source, runs entirely on the user's machine, and attaches to every agent already running there. https://github.com/umanitek/agent-blackbox
Here's why that matters. Picture an AI agent doing completely ordinary work: pulling up documentation, installing a package, reading a config file. Nothing to see. Then, without anyone telling it to, it reaches for a private SSH key. Nobody typed that command. A page it read earlier had hidden instructions buried inside it – invisible to the human, perfectly readable to the agent – that quietly rewrote what it did next. That's a prompt injection attack, and it's one of the sneakiest ways an AI agent can be turned against the person running it.
Blackbox is designed to catch this kind of behaviour. It starts in flag-only mode, monitoring and surfacing suspicious actions without blocking anything until the user decides to enforce protection.
The network compounds, creating a collective digital immune system. Threats one Blackbox identifies are anonymized and published to OriginTrail's Decentralized Knowledge Graph, a public, tamper-proof record no single party can rewrite. Every other Blackbox syncs against it. One agent's finding hardens every agent on the network.
Star the repo, read the docs, or install Blackbox in one command at: https://umanitek.ai/blackbox
