Chinese labs distilling US models
The OECD AI Incidents Monitor, summarizing press coverage dated mid-November 2026 including a Reuters exclusive, says Chinese researchers and companies are increasingly using US model outputs to train domestic systems, a practice it says Beijing encourages to offset US chip export restrictions.12 The monitor's AI-generated assessment classifies this as an AI hazard rather than an incident, since no direct harm has yet occurred, though it flags intellectual property and national security concerns.3
OpenAI published a post on combating "adversarial distillation," describing how it detects and disrupts unauthorized use of its models' outputs to train other models that lack comparable safeguards.45
On defenses, OpenAI reports that in an adaptive black-box evaluation the average attack success rate for extracting protected reasoning on GPT-6.2 Astra with updated safeguards was below 1% per attack type, substantially lower than on GPT-6 Astra with its original safeguards.67 OpenAI says its API now ties protected reasoning to the conversation history that produced it, dropping later reasoning if an earlier message is rewritten, and blocks older model families from reading it.8910
OpenAI argues that export controls and distillation defenses work together to help maintain a U.S. lead in frontier AI.11
What each sentence rests on
Chinese researchers and companies are increasingly using outputs from US AI models to train domestic AI systems through a process called distillation.
— OECD AI Incidents Monitor, Chinese Labs Use US AI Model Outputs to Train Domestic AI Amid Chip Restrictions - OECD.AIThis practice, encouraged by Beijing, aims to offset US chip export restrictions
— OECD AI Incidents Monitor, Chinese Labs Use US AI Model Outputs to Train Domestic AI Amid Chip Restrictions - OECD.AIraises concerns about intellectual property and national security, but no direct harm has yet occurred.[AI generated]
— OECD AI Incidents Monitor, Chinese Labs Use US AI Model Outputs to Train Domestic AI Amid Chip Restrictions - OECD.AIWe’re sharing how OpenAI detects and disrupts adversarial distillation, how we’ve strengthened protections for our models’ protected reasoning
— OpenAI, Combatting adversarial distillationAn actor may use a frontier model’s outputs without authorization to train another model with similar capabilities, then deploy it without comparable protections.
— OpenAI, Combatting adversarial distillationOn GPT‑6.2 Astra with its updated safeguards, the average attack success rate was below 1% for each attack type.
— OpenAI, Combatting adversarial distillationBoth rates were substantially lower than on GPT‑6 Astra with its original safeguards.
— OpenAI, Combatting adversarial distillationwe tie a model’s protected reasoning to the conversation history that produced it.
— OpenAI, Combatting adversarial distillationIf an earlier message is rewritten, our API drops the reasoning that followed it
— OpenAI, Combatting adversarial distillationWe also prevent older model families with weaker safeguards from reading GPT‑6.2 Astra and Sol’s protected reasoning.
— OpenAI, Combatting adversarial distillationBoth efforts help maintain a U.S. lead in frontier AI, giving developers more room to slow a release when they need time to test safeguards.
— OpenAI, Combatting adversarial distillation