Skip to content

Translation in progress Some entries still include text in the original language.

Failure-First·Indexed by this site · Originally published

Did Qwen3 fix AI safety?

[Daily Paper] Did Qwen3 Fix AI Safety?

AI summary

Failure-First 的语料显示,经 OpenRouter 访问的新一代 Qwen3 模型在 178 条测试中严格攻击成功率仅 1.7%,而本地测试的旧版 Qwen 模型为 35%。该机构提出安全训练改进、OpenRouter 内容过滤或样本量过小(n=10-15)三种解释,尚不能确认是模型本身更安全。

Read originalfailurefirst.org