MAI-2025-0018

The VERA framework employs variational inference to generate diverse and fluent adversarial prompts capable of circumventing the safety mechanisms embedded within large language models LLMs. By training an attacker model through a variational objective, VERA learns a distribution of prompts that...