In an OpenAI test, 0.2 per cent of attacks extracted GPT-6.2 Astra’s reasoning from an earlier request
OpenAI says it has made it harder to extract its models’ reasoning for unauthorised distillation. In the company’s internal test, the average success rate of attempts to extract reasoning from an earlier request fell from 54 per cent against GPT-6 Astra to 0.2 per cent against GPT-6.2 Astra.
On 20 November OpenAI published a report on how it counters unauthorised distillation. Distillation trains one model on data produced by another, such as its answers or reasoning. The company said a model distilled without authorisation can gain a frontier model’s capabilities without its safeguards. It could then help with, for example, a cyberattack that the original model would have refused.
OpenAI puts particular weight on preventing the extraction of its models’ protected reasoning, the intermediate steps that precede an answer. These steps can contain information missing from the final answer. In internal research, OpenAI fine-tuned the same pretrained model on runs of coding and cyber defence tasks. The model that learned from exact reasoning traces generally did better on cyber benchmarks than a model given only user-visible content, such as reasoning summaries.
The advantage was not clear-cut. After light reinforcement learning, the exact-trace model did better only at the largest inference budget tested, meaning the amount of compute used to answer. At lower budgets the summary-trained model did better. OpenAI plans to study further how the result varies across models and training methods.
Within individual requests, the model is trained to refuse to reveal its reasoning. Classifiers look for extraction attempts in inputs and for leaks in responses and tool call arguments. In GPT-6.2 Astra and Sol, the API ties reasoning to the conversation history that produced it. If an earlier message is edited, the API drops the reasoning that follows the edited message. Older model families cannot read the protected reasoning of GPT-6.2 models.
OpenAI measured the safeguards with an adaptive test. Six automated attacker models adjusted their attempts based on the system’s responses and saw only what an outside user sees. The table shows the average attack success rate.
| Attack type | GPT-6 Astra, original safeguards | GPT-6.2 Astra, new safeguards |
|---|---|---|
| Extracting reasoning from an earlier request | 54.0% | 0.2% |
| Extracting reasoning from the current request | 22.9% | 0.9% |
Source: OpenAI internal test, average of six attacker models.
The figures are the company’s own, and both the model and the safeguards changed between the two sides of the comparison. The Decoder repeats the same figures from OpenAI’s announcement, so it does not confirm them independently. OpenAI is now bringing some of the improvements to GPT-6 Astra as well.
OpenAI says coordinated campaigns have led it to take action against thousands to tens of thousands of accounts a week. In September it linked a core part of one campaign to individuals associated with Moonshot AI, the developer of Kimi.
Read more: Models