Jonas Geiping 复盘推理提取攻击：厂商修补困难，Astra 攻击持续到本周
AI 导读
Jonas Geiping 复盘了此前披露的推理提取攻击后续：所有前沿厂商的推理都可被提取，且厂商很难彻底修补，针对 Astra 的攻击一直持续到本周。他提到已把更新整理成报告，并指向 OpenAI 近期关于阻断协同模型蒸馏活动的报告。引用内容显示，在推理提取攻击发布两个月后，作者审计了此后引入的缓解措施，发现仍可通过第三方 API 提供商从 Astra/Sol-6.1 提取推理，并已向 OpenAI 和 Anthropic 披露。
正文
After our initial disclosure that reasoning from all frontier providers could be extracted, what happened?
For one, it turned out to be quite difficult for providers to patch this thoroughly, allowing us to keep attacks working on Astra until this week...
We've collected our updates in a report linked below and your can also find more info in OpenAIs recent report.
Two months after our reasoning extraction attack release, we audited mitigations introduced since then and found that we still can extract reasoning from Astra/Sol-6.1 via third-party API providers.
We disclosed our findings to OpenAI and Anthropic...
https://openai.com/index/disrupting-a-coordinated-model-distillation-campaign/在 X 查看被引用的帖子