Researchers say reasoning content can still be extracted from Astra/Sol-6.1 via third-party APIs
Two months after publishing the reasoning extraction attack, the authors audited the mitigations subsequently introduced by the vendor and found that reasoning content could still be extracted from Astra/Sol-6.1 through a third-party API provider, and they have disclosed the findings to OpenAI and Anthropic. The authors said the scale of the attack surface enterprises need to cover was surprising, and attacks may still bypass protections; the reasoning traces of Astra/Sol-6.1 are now open to the public, which is why the authors believe chain-of-thought monitoring may be quite difficult.
It was interesting to see how much attack surface exists for companies to cover and how we could still slip through with our attacks.
We now have Astra/Sol-6.1 traces for the public. I now see why CoT monitoring might be rough!
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/View the quoted post on X