Ironhide
I'm building Ironhide to test what AI agents do with their tools.
- Role
- Side project
- Status
- Currently building
- Year
- 2026
- Live
- ironhideai.com
- 2,000+
- Attack environments
- ~3.7k
- Tests
The problem
AI agents don't just answer questions any more. They take actions, call tools, move through environments, and hold real permissions. The blast radius of a mistake is no longer a bad paragraph. It's a write to a production system.
Application security tooling was built for software whose control flow you can read. It doesn't model an agent that decides what to do next at runtime, so the interesting failures, the ones that come from the agent reasoning its way into something it shouldn't do, go unexamined until they happen in production.
What I built
Ironhide runs a customer's agent against 2,000+ attack environments and grades the actual tool calls and state changes (not just the text) into PASS / FAIL / BLOCK verdicts, each with evidence and a one-command repro.
It lives in the developer workflow rather than beside it: a CLI gates pull requests in CI, the same way you'd run a test suite.
Selected details
- Where it sits
A CLI that runs in CI and gates pull requests, so verification happens on the path to production instead of in a separate security review.
- What it grades
The tool calls an agent makes and the state changes they cause, graded into PASS / FAIL / BLOCK with evidence and a one-command repro for every verdict.
- Stack
I'm building it with FastAPI and React, with roughly 3,700 tests.
Screens
