🛡️ OpenAI says copycats trying to steal its AI's "thinking" now succeed just 0.2% of the time
Rivals can build cheap clones by training on a top model's outputs, a trick called distillation. OpenAI says it has hardened GPT-6.2 Astra and Sol against attempts to pull out their hidden step-by-step reasoning: in internal red-team tests, attack success fell from 54% on GPT-6 Astra to just 0.2%. In its own study, a model trained on exact reasoning traces generally picked up more coding and cyber skill than one trained on summaries. The risk, OpenAI warns: clones copy the capabilities but not the safety guardrails.
⭐ The AI race has a new front: protecting how models think, not just what they answer.
#AI
Rivals can build cheap clones by training on a top model's outputs, a trick called distillation. OpenAI says it has hardened GPT-6.2 Astra and Sol against attempts to pull out their hidden step-by-step reasoning: in internal red-team tests, attack success fell from 54% on GPT-6 Astra to just 0.2%. In its own study, a model trained on exact reasoning traces generally picked up more coding and cyber skill than one trained on summaries. The risk, OpenAI warns: clones copy the capabilities but not the safety guardrails.
⭐ The AI race has a new front: protecting how models think, not just what they answer.
#AI
