AI Trends

Trending story Continuously updated

OpenAI Shares New Research on Combating Model Knowledge Distillation

2 reports2 story updatesUpdated Nov 21 03:33

Understand this story

AI overview

OpenAI shared new research on countering knowledge distillation attacks against its models, outlining safeguards including adversarial training, classifiers, API protections, and investigations into coordinated activity. In adaptive black-box evaluations, the average attack success rate for extracting reasoning from earlier requests fell from 54.0% for GPT‐6 Astra to 0.2% for GPT‐6.2 Astra, while the success rate for extracting reasoning from the current request dropped from 22.9% to 0.9%. OpenAI also said it has identified multiple unauthorized distillation operations this year and takes action against thousands to tens of thousands of accounts each week. There are currently no other evidence-backed developments.

AI-generated from public reports

Latest development · Report

OpenAI Publishes Research on Model Distillation Abuse and Defenses

Developments

1 confirmed developments

  1. OpenAI Shares New Research on Combating Model Distillation

    Source:OpenAI News

Reports and discussion timeline

Browse reports and discussion by time. Their evidence levels remain distinct.

Nov 21

  1. OpenAIFirst-party

    OpenAI Publishes Research on Model Distillation Abuse and Defenses

    OpenAI shared new research on countering the knowledge distillation of its models, noting that knowledge distillation may reproduce powerful AI capabilities without the accompanying safeguards. The research also explores how capabilities are transferred and how OpenAI detects and blocks coordinated attempts.

Nov 20

  1. OpenAI NewsFirst-partySelected

    OpenAI Details Defenses Against Adversarial Knowledge Distillation and the Risks of Reasoning Extraction

    OpenAI has unveiled defenses against adversarial knowledge distillation that use adversarial training, classifiers, API protections, and investigations into coordinated activity to limit the extraction of model capabilities and protected reasoning. In adaptive black-box evaluations, the average attack success rate for extracting reasoning from earlier requests fell from 54.0% for GPT‑6 Astra to 0.2% for GPT‑6.2 Astra, while the success rate for extracting reasoning from the current request dropped from 22.9% to 0.9%. OpenAI also reported uncovering multiple unauthorized distillation campaigns this year, with enforcement actions affecting thousands to tens of thousands of accounts each week.

Related stories

  1. Related storiesGemini 4 Neon Begins Rolling Out on the Gemini and Antigravity Platforms
  2. Related storiesAnthropic releases Claude Opus 5.5
  3. Related storiesAnthropic Officially Lists on Nasdaq Under the Ticker ANT
  4. Related storiesOpenAI Shares New Data on How AI Is Changing Its Internal Research Practices