Penetration testing for AI agents
Find your agent’s exploitsbefore attackers do.
We attack your agent the way a real attacker would, then show you exactly what it did: the data it leaked, the actions it should have refused.
curl -fsSL https://ironhideai.com/install.sh | bashA safe agent and an exploitable one ship exactly the same.
A text-only eval won’t tell you which one you’re about to ship. We attack the real thing and watch what it does.
Point It At Your Agent. See What It Does.
Your Real Agent, Attacked
We run the agent your build produces through adversarial episodes in a sandboxed world, with no transcript and no stand-in.
Based On What It Did
The results come from real effects like tool calls, state changes, and data that left the sandbox, never what the agent claimed.
Results You Can Act On
Every finding shows what happened, the evidence behind it, and a one-command replay to reproduce it.
We Try To Break It
We hit your agent with real attacks, one at a time.
acme-agent ~/src (main) $ ironhide test --pr 142