arXivDaily arXiv每日学术速递 周一至周五更新
arXiv 2602.16729cs.CRcs.AIcs.CLcs.LG
收藏

意图洗白:AI安全数据集并非看起来那样

Intent Laundering: AI Safety Datasets Are Not What They Seem

Applied Machine Learning Research(应用机器学习研究)

Shahriar Golchin, Marc Wetter

AI总结:

研究评估了广泛使用的对抗安全数据集质量,发现其过度依赖触发线索,导致模型安全评估与真实攻击行为脱节。

摘要:

我们系统地评估了广泛使用的对抗安全数据集的质量,从孤立和实际两个角度进行考察。在孤立情况下,我们检查这些数据集是否能反映真实世界的对抗攻击,基于三个定义属性:由隐含意图驱动、精心设计、非分布。我们发现这些数据集过度依赖“触发线索”:具有明显负面/敏感含义的词或短语,旨在显式触发安全机制,这与真实世界的攻击相比是不现实的。在实际情况下,我们评估这些数据集是否真正衡量安全风险,还是仅仅通过触发线索引发拒绝。为此,我们引入了“意图洗白”:一种从对抗攻击(数据点)中抽象出触发线索的程序,同时严格保持其恶意意图和所有相关细节。我们的结果表明,由于过度依赖触发线索,当前对抗安全数据集无法真实反映现实中的对抗行为。一旦这些线索被移除,所有之前评估为“合理安全”的模型都变得不安全,包括Gemini 3 Pro和Claude Sonnet 3.7/4。此外,当意图洗白被适合作为绕过技术时,它在完全黑盒访问下始终能获得高攻击成功率,范围从90.00%到100.00%。总体而言,我们的发现揭示了现有数据集评估模型安全性和现实世界对手行为之间存在显著脱节。

英文摘要:

We systematically evaluate the quality of widely used adversarial safety datasets from two perspectives: in isolation and in practice. In isolation, we examine how well these datasets reflect real-world adversarial attacks based on three defining properties: being driven by ulterior intent, well-crafted, and out-of-distribution. We find that these datasets overrely on "triggering cues": words or phrases with overt negative/sensitive connotations that are intended to trigger safety mechanisms explicitly, which is unrealistic compared to real-world attacks. In practice, we evaluate whether these datasets genuinely measure safety risks or merely provoke refusals through triggering cues. To explore this, we introduce "intent laundering": a procedure that abstracts away triggering cues from adversarial attacks (data points) while strictly preserving their malicious intent and all relevant details. Our results show that current adversarial safety datasets fail to faithfully represent real-world adversarial behavior due to their overreliance on triggering cues. Once these cues are removed, all previously evaluated "reasonably safe" models become unsafe, including Gemini 3 Pro and Claude Sonnet 3.7/4. Moreover, when intent laundering is adapted as a jailbreaking technique, it consistently achieves high attack success rates, ranging from 90.00% to 100.00%, under fully black-box access. Overall, our findings expose a significant disconnect between how existing datasets evaluate model safety and how real-world adversaries behave.

机构、总结、中文摘要均由AI分析而来,可能存在错误,请以论文正文为准。

↑