OpenAI Reveals Progress on Anti-Distillation Protections: GPT-6.2 Astra Chain-of-Thought Extraction Success Rate Falls from 54% to 0.2%
IT Home, November 21 — On November 20 local time, OpenAI published an announcement detailing its defenses against "Adversarial Distillation."
OpenAI said it has strengthened protections for the internal reasoning Chain-of-Thought of GPT-6.2 Astra and GPT-6.2 Sol, in order to reduce the risk of other models acquiring their capabilities by extracting the reasoning process.

OpenAI stated that adversarial distillation refers to the unauthorized use of the outputs of frontier AI models to train other models, causing the latter to acquire similar capabilities without the corresponding safety protections of the original models. Such behavior could lead to high-risk capabilities in areas such as cybersecurity and biology being transferred to models with weaker safety restrictions.

In internal adaptive black-box red-team testing, OpenAI tested the success rate of attackers extracting protected reasoning content from a model's previous requests and current requests.
The results showed that the extraction success rate for reasoning content from previous requests fell from 54.0% under GPT-6 Astra's original safety mechanisms to 0.2% after GPT-6.2 Astra's updated safety mechanisms; the extraction success rate for reasoning content from current requests fell from 22.9% to 0.9%. The test results aggregated data from 6 attack models.
To achieve these protections, OpenAI adopted multiple layers of measures, including adversarial training that enables models to recognize and refuse reasoning-chain extraction requests, input and output classifiers to detect relevant attacks and reasoning content leakage, and API restrictions on the reuse of reasoning content across different sessions, models, and organizations. IT Home's accompanying image is as follows:

Among these, GPT-6.2 Astra and GPT-6.2 Sol associate protected reasoning content with the conversation history that generated it. If earlier messages are modified, the API discards subsequent reasoning content generated based on that message. In addition, older models with weaker safety protections cannot read the protected reasoning content of these two new models.
Regarding enforcement against related behavior, OpenAI said that since the start of this year, the company has taken action every week against accounts involved in unauthorized distillation activities, ranging in scale from thousands to tens of thousands, with enforcement clearly increasing since this summer. The company also investigates networks of accounts coordinating large-scale extraction and shares attack indicators with other model providers to help identify related abuse.
OpenAI also described its research on the safety risks of distillation. Internal testing found that, compared with student models trained only on the model's final answers, reasoning summaries, and tool interaction records, student models supervised-fine-tuned on the teacher model's complete reasoning traces generally achieve better transfer of cybersecurity capabilities.
- "Fields Medal winner Deng Yu comments on AI companies' anti-distillation: You can't support distillation only when you're distilling all of humanity's knowledge"
- "NVIDIA CEO Jensen Huang responds to AI model distillation controversy: This is market competition behavior"
- "Commerce Ministry responds to U.S. so-called AI distillation accusations: No evidence, no legal basis, a typical double standard"
- "Report: Zhang Yiming issues internal strict order: Even if the model temporarily lags behind, ByteDance will not rely on AI distillation technology"
- "Microsoft CEO Nadella criticizes AI model companies' 'double standards': Advocating 'fair use' of data while restricting others from distilling"
Related Articles
- OpenAI launches MedChemBench benchmark to evaluate AI-assisted drug R&D capabilities
- OpenAI announces Spaces collaborative workspace open to ChatGPT Plus users, supporting joint editing by multiple people
- OpenAI invites some Plus users to try the dots AI agent; mobile ChatGPT will give it a dedicated tab
- OpenAI opens GPT-Live-1.1 API: supports real-time two-way conversation, agents can chat while handling complex tasks
- Apple is developing its own AI agent: can control devices, browse the web, and write code, aiming to compete with OpenAI and Anthropic
- ChatGPT adds "real-time dictation" feature: speak into the microphone and speech turns into text instantly