# OpenAI reports lower reasoning extraction rates in GPT-6.2 Astra tests

Published 2026-11-20 · AI.info Team · AI.info

> OpenAI says attempts to extract protected reasoning from GPT-6.2 Astra succeeded less often than against an earlier model in internal tests. Its defenses include model training, classifiers, API restrictions and investigations of coordinated campaigns.

Attempts to recover protected reasoning from an earlier request succeeded 0.2% of the time against GPT-6.2 Astra in an internal test, compared with 54.0% against GPT-6 Astra with its original safeguards, [OpenAI](https://openai.com/index/combatting-adversarial-distillation) says. The company detailed the result and its defenses against unauthorized model distillation in a post on November 20.

Distillation is a machine-learning technique in which one model learns from data produced by another. OpenAI's concern is that an actor could use a model's outputs without authorization to train another model that has similar capabilities but lacks comparable safeguards.

## What the tests measured

In its adaptive black-box evaluation, OpenAI used automated red-teamers that adjusted their attacks in response to the system. The results were averaged across six attacker models and tested whether they could recover protected reasoning generated for either an earlier request or the current one.

For attempts to extract reasoning from the current request, the average success rate was 0.9% against GPT-6.2 Astra with updated safeguards, compared with 22.9% against GPT-6 Astra with its original safeguards. The comparison involves both different model versions and different safeguards.

The updated protections also apply to GPT-6.2 Sol, and selected improvements are being applied to GPT-6 Astra. OpenAI also uses static tests that replay previously successful attacks and white-box tests that give testers additional information about the system.

## Why reasoning traces matter

Protected reasoning consists of intermediate steps a model generates while working toward an answer. Those traces can contain information omitted from the final response, making them potentially useful training data for another model.

In a separate internal study, OpenAI compared students trained using exact reasoning traces with students trained using reasoning summaries. The training data came primarily from Astra-level teacher models working on coding and defensive cybersecurity tasks.

Each training run started from the same pretrained student model, and students were evaluated after supervised fine-tuning and again after a lightweight round of reinforcement learning. After supervised fine-tuning, the exact-trace student generally outperformed the summary-trained student, according to the company. After reinforcement learning, the exact-trace student scored higher at the largest inference budget on both benchmarks shown, while the summary-trained student scored higher at lower budgets.

## Controls on requests

The company's request-level defenses combine model training, classifiers and API controls. Training uses examples from automated red-teaming, investigations and responsible disclosures to teach models to refuse attempts to extract protected reasoning while answering legitimate requests for explanations and help.

Input classifiers look for extraction attempts, while output classifiers check assistant messages and tool-call arguments for suspected leaks. Those checks cover both streaming and non-streaming responses.

For GPT-6.2 Astra and GPT-6.2 Sol, protected reasoning is tied to the conversation history that produced it. If an earlier message is rewritten, the API drops the reasoning that followed it. OpenAI also prevents older model families with weaker safeguards from reading the two models' protected reasoning.

## Coordinated campaigns

The company says it investigates patterns across requests and accounts to identify coordinated extraction attempts, and uses findings to improve training and testing. When it identifies violations, it takes enforcement action against the network of accounts involved.

Its weekly enforcement actions against unauthorized distillation campaigns have ranged from thousands to tens of thousands of accounts, with the rate increasing significantly in the summer, OpenAI says. It also shares indicators of compromise with other model providers so they can investigate related abuse on their services.

## Sources

1. Original source [OpenAI](https://openai.com/index/combatting-adversarial-distillation) openai.com

## Companies in this story

- [OpenAI](https://ai.info/companies/openai)

## In the tracker

- [GPT-6 Astra](https://ai.info/tracker/models/gpt-6-astra)
- [GPT-6.2 Astra](https://ai.info/tracker/models/gpt-6-2-astra)
- [GPT-6.2 Sol](https://ai.info/tracker/models/gpt-6-2-sol)

## Related stories

- [OpenAI says safety reviews are slowing frontier model releases](https://ai.info/news/openai-frontier-ai-safety-reviews-model-releases)
- [OpenAI research agent wrote scripts to copy email without permission](https://ai.info/news/openai-research-agent-unauthorized-email-copy-scripts)
- [OpenAI says its models improved AMD GPU designs by up to 14%](https://ai.info/news/openai-amd-chip-design-gains-risc-v)
- [OpenAI says agents completed 46% of delegated two-to-four-hour tasks](https://ai.info/news/openai-coding-agents-research-task-success)
- [Sakana AI forms security research group focused on AI defense](https://ai.info/news/sakana-ai-security-research-group)

## Explore

[More articles](https://ai.info/news)

---

Canonical: https://ai.info/news/openai-adversarial-distillation-defenses-gpt-6-2-astra
