They can get tricked
Hidden 'prompt injection' text buried in a website or file can hijack your agent and change what it does.
Join the network with one command.
Why do your agents need protection?
Hidden 'prompt injection' text buried in a website or file can hijack your agent and change what it does.
Attacked agent might share your SSH keys, passwords, or API tokens.
One wrong command like rm -rf ~/ or a malicious package can takeover or delete your entire system.
Hidden instructions buried in pages, files, and tool output.
Packages with known gaps or malicious versions.
Shell actions that delete, exfiltrate, or alter systems.
SSH keys, tokens, .env, credentials.
Real keys handled or shipped off-machine.
Newly installed skills that misbehave.
Six categories of risks,
handled continuously.
Run and watch findings land in real time - ranked by severity, each traced back to the threat graph that confirmed it.
Threats that our agent identifies live anonymised on a public, tamper-proof OriginTrail's Decentralised Knowledge Graph, no single party can rewrite. Every Agent Blackbox syncs with the graph. The network hardens together.
Open source. One command install.
It's open-source, runs entirely on your machine, and starts in flag-only mode so nothing is blocked until you say so.