3 ms·
The regulator isn't taking user input directly, but it is being exposed to output from a model that was created in response to user input. So the adversarial a
by simonw 4y ago
The regulator isn't taking user input directly, but it is being exposed to output from a model that was created in response to user input.
So the adversarial attacker needs to come up with original input that will produce output that itself contains a successful prompt injection attack targeting the regulator.
Harder, but I doubt impossible.