SpiralAI To Restrict User Input In RyzaChat:AI After Prompt Injection Attempts
The policy shifts from filtering what Ryza says to limiting what players can type, an admission that the game master layer and refusal behavior were not enough to hold the character's boundaries once users started probing them.
Reporting from 1 source: Automaton.
SpiralAI said on September 29 that it will change the specifications of RyzaChat:AI, its licensed AI chat RPG set in the Atelier Ryza world, to restrict what users can enter. The company cited prompt injection attempts by some users and said it does not tolerate attempts to make characters say things contrary to public order and morals. An AI model update is planned for October.
Existing safeguards included Ryza declining topics outside her worldview, a game master AI that breaks down user input before passing it to the character, and a stop on inappropriate expressions just before display. Content contrary to public order and morals is also processed so it does not enter the character's memory. SpiralAI said some users have tried prompt injection to pull out responses that should be restricted anyway.
Alongside the October model update, the company will limit user input content and later consider adjusting the character limit based on how things go. It has also been filing deletion requests over altered or fabricated images and videos presented as screens from the app, and repeated its August 27 call to play within the terms of service.
Synthesized by Yomimono from the 1 cited source below, including Japanese-language reporting where cited, then editorially reviewed before publishing.