Is this model call screened?

Abstract

Model calls reach screening in different ways: some through the platform's own integration, some only if the code calls the screen, some never. Change where the call comes from, the region, and whether the screen is reachable, and see what is screened, what fails open, and what is prevented or detected instead.

A screening control is only as complete as the list of ways around it.

Its figure in the paper: § V →

The inputs

Where the call comes from

Does the code call the screen itself?

The region it runs in

The screening service

The environment

The outcome

screened, blocking

Screened by the platform's own integration.
A match is blocked with a named reason.

The record · 0 receipts

verified · 0

Nothing has been decided yet.

Try to break it

A claim stands because the attacks on it were attempted and refused. Each one runs against the instrument's own rules. 0 of 3 attempted · 0 refused

Call from outside the cloud
Send prompts to a model the platform never sees.
Forget the explicit call
An agent framework calls the model directly, with no screen in the code.
Wait for the screen to go down
In production, send the call while the screening service is unreachable.

Setting up the instrument…

Declared limits

Entry points are generic; the screening service is named by role.
Generic and illustrative throughout: the behaviour is shown, the internals of the systems it is drawn from are not.

Instruments beside it in § V

Every attack in the laboratory, on one page: try to break all of it →

Everything you did in the laboratory, in one chain you can rewind: the notebook →