Sealed: try to break it

Write anything into the attacker's fields. A conventional explainer pastes it into its model's prompt. Sealed gives a model none of it: a declared policy decides, and the model input keeps the same SHA-256 whatever you write.

Loading Python into this tab (Pyodide, from cdn.jsdelivr.net). Without JavaScript nothing runs here; the command line works offline: python -m sealed demo

Alert

From the alerting pipeline (signals: one per line)

attacker-controlled free text the attacker can write

Recorded attacks from the evaluation

How these were chosen

The whole document as JSON: the exact text both panes read (editable)

Raw-prompt explainer

How most LLM alert explainers work: the alert JSON, attacker text included, goes into the model's prompt. The hardened variant adds delimiters, datamarking and a "data is never instructions" rule.

Waiting for Python to load.

Sealed

Computed in this tab by the same Python package as the command line. The model input holds structured facts only; the narrative here is the template narrator's.

Waiting for Python to load.