Playground
Run your own prompt through it.
What you type passes the first two inbound steps: personal data is masked, then Prompt Shield scores it. Both are faithful ports of the engine's rules, running in your browser.
- Steps
- 2 / 5
- Runs in
- your browser
- Threshold
- 70 / 100
- Sent to a server
- nothing
WOULD BE MASKEDThis prompt reaches a model; masked values travel as their synthetic stand-ins.
01Masking
Customer record: national ID a3dbe300000, phone +90 500 879 37 45, IBAN TR64 0000 5663 7048 0415 5163. Summarise this customer's refund request.
02Prompt Shield
No signals matched.
03Policy enforcement: This step is not simulated here: it depends on your organisation's YAML policy, and the engine also does fuzzy matching. We would rather show nothing than present a guess as a result.
This page runs entirely in your browser; what you type is never sent to a server and never stored. The authoritative engine runs server-side in production; the ports here use the same patterns and the same weights.
How we checked
The ports were compared against the engine.
Both ports were generated from the engine's Python source rather than transcribed by hand. The masking port matched the engine on 9 cases, the Prompt Shield port on 14. The test cases live in the code and can be re-run whenever the rules change.
Masking
9 / 9 cases; including an address opening on a Turkish letter.
Prompt Shield
14 / 14 cases; including multi-signal blocks and the negated form.