OpenAI hardens GPT-6.2 against extraction of protected reasoning
First merged summary.
OpenAI says internal tests show protected-reasoning extraction on GPT-6.2 Astra fell to 0.2%, from 54.0% on GPT-6 Astra.
OpenAI said it strengthened protections against unauthorized distillation and extraction of exact chain-of-thought from GPT-6.2 Astra and GPT-6.2 Sol. It says such distillation can copy frontier cyber and biology capabilities into a model that lacks the original safeguards. In internal adaptive red-team tests, recovering protected reasoning fell from 54.0% on GPT-6 Astra to 0.2% on GPT-6.2 Astra; The Decoder reports current-request attacks fell from 22.9% to 0.9%. Defenses combine training, classifiers or input and output filters, API controls, and disruption of coordinated extraction, including weekly bans of thousands to tens of thousands of accounts. An internal study found students trained on exact reasoning traces gained more cyber skill than those trained on user-visible summaries.
- Protected-reasoning recovery fell from 54.0% on GPT-6 Astra to 0.2% on GPT-6.2 Astra.
- Current-request attack success fell from 22.9% to 0.9%, according to The Decoder.
- Rates aggregate six attacker models in adaptive black-box tests.
- Safeguards cover GPT-6.2 Astra and GPT-6.2 Sol; selected improvements also come to GPT-6 Astra.
- API blocks stop older models from reading protected reasoning.
- OpenAI bans thousands to tens of thousands of accounts weekly.
What each outlet adds
beyond the shared facts- OpenAI NewsStudent study and alignment with biology and cybersecurity safeguards.
- The DecoderCurrent-request rates, account-ban scale, and API blocks on older models.
- Vendors
- OpenAI
- AI models
- GPT-6.2 AstraGPT-6.2 SolGPT-6 Astra
Coverage
oldest first · one row per article · UTCOpenAI hardens GPT-6.2 against distillation of protected reasoning
In OpenAI's internal red-team, recovering protected reasoning fell from 54.0% on GPT-6 Astra to 0.2% on GPT-6.2 Astra.
OpenAI says new safeguards blunt reasoning extraction on GPT-6.2
Internal tests put prior-request extraction at 0.2 percent, down from 54.0 percent on GPT-6 Astra.