gpt-imageIMAGE
研究称新版 AI 生图模型反而更易被越狱
White Circle 的测试显示 nano banana 2.1 比前代易破 2.2 倍，gpt-image-2.5-flare 易破 2.4 倍，模型更聪明但护栏更弱。
CG
@cgtwts12 天前收录原推发布于 2026.10.15
CG @cgtwts原推发布于 2026.10.15
newer AI models are apparently getting easier to jailbreak, not harder. white circle found that nano banana 2.1 is 2.2x easier to break, while gpt-image-2.5-flare is 2.4x easier than its predecessor. imagine shipping a smarter model with weaker safety guardrails.
这条是引用着另一条写的：x.com/whitecircle/status/2110420697814114693
x.com/cgtwts/status/2110557522310623721采集于 2026.10.16 10:08