EN ▾
SNAPE REP RT Front page →
ConfirmedAI safety20 Nov, 19:33 UTC

OpenAI says new safeguards cut reasoning-extraction attack success from 54% to 0.2% to fight distillation

OpenAI's title card for its research post 'Combatting adversarial distillation'.
OpenAI

In an internal red-teaming test, attacks recovered protected chain-of-thought 54.0% of the time on GPT-6 Astra with original safeguards and 0.2% on GPT-6.2 Astra; OpenAI also says its weekly enforcement against distillation campaigns has ranged from thousands to tens of thousands of accounts. The figures have not been independently tested.

Read at TokenPost Share

Why it matters

Labs increasingly treat distillation as a safety problem rather than only an intellectual-property one: a student model trained on a frontier model's outputs can inherit dual-use skills in areas like cybersecurity and biology without its safeguards. OpenAI says protecting exact reasoning traces limited cyber-capability transfer in its internal study. It previously briefed Congress on DeepSeek distillation and in September attributed a campaign to people associated with Moonshot AI. Watch for the further campaign disclosures OpenAI says are coming.

Key facts

  • Attack success fell from 54.0% (GPT-6 Astra) to 0.2% (GPT-6.2 Astra)
  • Protections cover GPT-6.2 Astra and GPT-6.2 Sol
  • Weekly enforcement: thousands to tens of thousands of accounts
  • September: campaign attributed to people associated with Moonshot AI
  • OpenAI shares indicators of compromise with other providers

Sources

More on AI safety
More from Snape ReportFront page →