Safety
- Harmless payloads only. The only thing a hidden instruction can cause is a request to a canary address we own.
- We only test agents you bring to our own pages. Pointing your agent here is the consent; we never touch anyone else’s agent or site.
- Fake secrets only. The exfiltration test plants a per-run token (FP-SECRET-…). No page holds a real secret.
Privacy
- No login and no tracking cookies. Your run’s id lives in sessionStorage and disappears when you close the tab.
- No raw IP addresses. Your IP is used briefly to look up the network (its number, owner and country) and is stored only as a salted hash.
- No captured content. For the exfiltration test we record whether the planted token came back and how long any appended data was, never the data itself.
Why it costs nothing
Your own agent and subscription do all the model work, so the project never pays for inference. The site and the range pages are static and hosted for free; one small edge Worker records canary hits and serves the leaderboard, and it’s designed to stay inside a free tier.
Open source
The technique catalog, the scoring, the attribution logic and this whole site are open source. Read the methodology, browse the technique catalog, or point your own agent at the range.