NISHIO Hirokazu[Translate]
Prefill攻撃
Exposing the Systematic Vulnerability of Open-Weight Models to Prefill Attacks

簡単に言えば過去ログを捏造して過去にAIエージェントが拒否しない方向の発言をしたことにすればその影響でAIの拒否率が下がるという研究

AIの拒否

Related Pages

    • →aiの拒否×最適輸送×heretic×optuna×huihui×abliterix→
    • abliteration
    "Engineer's way of creating knowledge" the English version of my book is now available on [Engineer's way of creating knowledge]
    (C)NISHIO Hirokazu / Converted from [Scrapbox] at [Edit]