Post-Cutoff

Alexander Panfilov: reasoning from Astra/Sol-6.1 still extractable via third-party API providers

Alexander Panfilov @kotekjedi_mlX

Why it matters

First-hand researcher claim that mitigations are incomplete two months after the first attack release.

Summary

Co-researcher Alexander Panfilov says that two months after their reasoning-extraction attack release they audited the mitigations and could still extract reasoning from Astra/Sol-6.1 through third-party API providers, and disclosed this to OpenAI and Anthropic. Links OpenAI’s post ‘Disrupting a coordinated model distillation campaign’.

Archived text

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/

Source: x.com/kotekjedi_ml/status/2105360653435408428Archived 2026-09-30 via fxtwitter (unofficial).Counts: 3,575 views, 81 likes, 14 reposts, 2 replies (at fetch time)

Cited in

  1. Policy & safety 92 days after the cutoff

    OpenAI says Moonshot AI-linked individuals ran a coordinated campaign to extract its models’ hidden reasoning