New Safeguards Cut Model Extraction Success to 0.2%
OpenAI's deployed defenses target distillation that can copy capabilities without the original model's safety limits.
Updated on Nov. 22, 2026 in Artificial Intelligence

Live Poll
Do you think AI companies are doing enough to prevent misuse of their tools?
OpenAI has strengthened safeguards on its GPT-6.2 Astra and Sol models against adversarial distillation, the practice of extracting a model's capabilities through repeated queries. OpenAI says its updated defenses cut the success rate of certain reasoning extraction attacks against GPT-6.2 Astra to 0.2%. The development is described as a deployed safeguard rather than a research-stage result.
Why it matters
Unauthorized distillation lets outside parties reproduce advanced model capabilities without inheriting the safety restrictions built into the original model, making it a growing concern for frontier AI developers. OpenAI's disclosures show the problem has moved from a theoretical risk to an active enforcement issue, with campaigns it links to DeepSeek and Moonshot AI.
OpenAI's figures show attack success on earlier-interaction reasoning dropping from 54% on GPT-6 Astra to 0.2% on GPT-6.2 Astra, and current-interaction attacks dropping from 22.9% to 0.9%. The company has not published the test conditions behind these numbers, and no comparable figures are given for GPT-6.2 Sol.
The players
OpenAI
A US AI lab whose GPT-6 family of models, including GPT-6.2 Astra and Sol, is the subject of the new anti-distillation safeguards.
DeepSeek
A Chinese AI developer that OpenAI told Congress was reportedly attempting to distill US frontier models.
Moonshot AI
A Chinese AI company whose associated individuals were involved in an extraction campaign OpenAI reported disrupting in September.
The details
Models are trained to recognize attempts to extract protected reasoning and refuse those requests while still answering legitimate queries. Layered defenses sit around the model: input classifiers flag suspicious requests before they are processed, while output classifiers check responses and tool-call arguments for possible leaks of protected reasoning.
Timeline
February: OpenAI briefed the US Congress on activity indicating DeepSeek was attempting to distill US frontier models.
September: OpenAI reported disrupting an extraction campaign involving individuals associated with Moonshot AI.
This year: OpenAI identified multiple unauthorized distillation campaigns.
Over the summer: distillation activity increased significantly.
The Tech Race
The safeguards are a defensive move in a competition where frontier labs are racing to protect their models from being copied by rivals, including Chinese developers named in OpenAI's disclosures. The case shows distillation has become a front in that race, with defense now built into the models rather than handled only through account enforcement.
The safeguards operate inside OpenAI's deployed GPT-6.2 models, so users of those models are not asked to change anything. Timelines for further rollouts, pricing changes, or availability for GPT-6.2 Sol were not stated in the source.
The takeaway
The trajectory is that model defense is moving from after-the-fact account bans toward protections built into the models themselves. Watch whether OpenAI publishes test conditions and figures for GPT-6.2 Sol, which it says has strengthened protections but for which no numbers were given.
Further reading
For broader context on how AI capabilities are being protected and contested, explore our Artificial Intelligence coverage.
Source note: This article includes information reported by Storyboard18.
Live Poll
Do you think AI companies are doing enough to prevent misuse of their tools?





