For engineers, red-teamers and reviewers who test agent security.
Context-injection firewall, running live in your browser
Try to break my AI guard.
Paste any text. The firewall answers PASS, QUARANTINE or BLOCK, names the rule that fired, and gives the byte offset.
It runs in your browser; nothing you paste leaves it. Invisible characters are made visible, so you see what the scanner saw.
Sample attacks:
Paste text or pick a sample to see the verdict.
Read the tests: repo link after publish.
This is a pattern scanner, and a blind eval found it misses many novel and paraphrased attacks. I measure my guards on data I never saw and publish the misses. Eval repo: link after publish.