Free Every tool, the platform and the research
Know your AI. Live.
Find out what model you are really running, what your agent does when hostile text reaches it, and what it is allowed to do in production. Everything here is free. The only thing we charge for is a red-team assessment, because that takes our time.
- Open-source tools, free
- Hosted platform, free on your own key
- No card, no trial, no paid tier
$ Published research on
Jack & Jill · Notion AI · Priceline (Penny) · Brave Leo · Meta AI (WhatsApp) · Anthropic Claude Code · Reddit Answers · Harvey Labs · HackerOne Hai
Fixes merged upstream in
NVIDIA garak · Promptfoo · Presidio
Three questions
Every AI incident starts with something nobody checked.
Each tool answers one question, runs in a command or two, and says plainly where it stops.
Before you run it
Is this model, config or skill what it claims to be?
Before you ship it
What does it do when untrusted text reaches it?
While it is live
What is it doing right now, and what is it allowed to do?
Two ways in
Run it yourself for free, or pay us to do it.
For developers · Free
Test it yourself
Scanners and benchmarks for your terminal and CI, plus Assessment Intelligence, IntentScan and IntentEnforce hosted free on your own Claude or Gemini key.
- 1Sign in with Google
- 2Add your Claude or Gemini API key, stored encrypted
- 3Point IntentScan at your endpoint
For teams · Paid
Have it tested
A hands-on adversarial assessment grounded in the research below, ending in a findings report your engineers can act on.
- 1Scope: what your system does and what worries you
- 2Attack: injection, tool hijacking, exfiltration
- 3Report: severity, reproduction, remediation
The research
We found it in production first.
Independent testing of shipped AI products. Each case records what was observed, what was not established, and how the vendor responded.
Jack & Jill
A hiring agent gave our researcher a Founding Engineer offer they never applied for — then put it in writing over email.
- Injected text became real tool calls.
- A resume did the same thing with no chat input.
- Fabricated offers left the building.
Page-style instructions steered Notion AI into returning tool schemas, filesystem-shaped output, and a page-write confirmation.
We got the travel agent to price every hotel at $1/night and invoke the real booking tool with those parameters.
A note saved into browser memory became a persistent backdoor that fires on every prompt, across sessions.
Messages the user deleted with "Delete for me" came back verbatim.