OpenAI Strengthens Defenses Against Model Distillation

Updated safeguards cut the success rate of reasoning extraction attacks on GPT-6.2 Astra to 0.2%, according to OpenAI.

Updated on Nov. 22, 2026 in Artificial Intelligence

Wide photograph of a long aisle of dark server racks with overhead fiber-optic cables curving toward a distant vanishing point under cool blue light.
OpenAI said its updated safeguards on GPT-6.2 Astra cut the success rate of certain reasoning-extraction attacks to 0.2%, as unauthorized distillation campaigns grow. AI Illustration. Upload story photo >

Live Poll

Do you think AI companies are doing enough to prevent misuse of their tools?

OpenAI has strengthened safeguards on GPT-6.2 Astra and Sol against adversarial distillation, reducing the success rate of certain reasoning extraction attacks against Astra to 0.2%. The company says its models are trained to recognize and refuse attempts to extract protected reasoning while still answering legitimate queries.

Why it matters

Unauthorized distillation can let attackers reproduce advanced model capabilities without the safety restrictions built into the original model. OpenAI says it has identified multiple such campaigns this year, with weekly enforcement actions affecting thousands to tens of thousands of accounts.

Attack success on earlier-interaction reasoning dropped from 54% on GPT-6 Astra to 0.2% on GPT-6.2 Astra, while current-interaction attacks fell from 22.9% to 0.9%. OpenAI has not stated how these figures were independently validated.

The players

OpenAI

OpenAI is an artificial intelligence research and deployment company that develops the GPT family of large language models.

DeepSeek

DeepSeek is a Chinese artificial intelligence company that develops large language models and has drawn scrutiny from US officials over alleged model distillation.

Moonshot AI

Moonshot AI is a Chinese artificial intelligence startup that develops large language models and was linked by OpenAI to an extraction campaign.

The details

Models are trained to recognize attempts to extract protected reasoning and refuse those requests. Input classifiers flag suspicious requests, and output classifiers check responses and tool-call arguments for possible leaks.

Timeline

  1. February: OpenAI briefed the US Congress about activity indicating DeepSeek was attempting to distill US frontier models.

  2. September: OpenAI reported disrupting an extraction campaign involving individuals associated with Moonshot AI.

  3. This year: OpenAI identified multiple unauthorized distillation campaigns.

The Tech Race

OpenAI's disclosure follows its February briefing to US Congress on suspected DeepSeek distillation of US frontier models. The company's defenses now target the same extraction techniques used by rival developers in competition over frontier AI capabilities.

Users of OpenAI's GPT-6.2 models are less likely to have protected reasoning extracted by third parties, which the company says preserves the safety restrictions built into the models. Developers building on these models may see stricter refusals on requests that resemble reasoning extraction.

The takeaway

Distillation attacks can copy advanced AI capabilities without the safeguards that came with them, so OpenAI's reported drop in attack success is a useful signal for anyone evaluating how well a model's protections hold up.

Further reading

For more on the AI race between US and Chinese developers, see Artificial Intelligence.

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?