Researchers re-test reasoning extraction attack: reasoning content can still be extracted from Astra/Sol-6.1 via third-party APIs two months later

Two months after the reasoning extraction attack was made public, the authors audited the mitigations introduced since then and found that reasoning content could still be extracted from Astra/Sol-6.1 through a third-party API provider. The authors have disclosed the findings to OpenAI and Anthropic, and included a link to OpenAI's announcement about cracking down on coordinated model distillation activity.