Skip to content
Brief Security & Trust ·

OpenAI: attacks to recover earlier reasoning on GPT-6.2 Astra fell from 54.0% to 0.2%

The figures come from OpenAI's own adaptive black-box test of attacks on one model's hidden reasoning.

In brief
  • In OpenAI's own test, attacks recovering earlier reasoning averaged 54.0% success on GPT-6 Astra and 0.2% on GPT-6.2 Astra. Attacks on the current request were below 1%.
  • OpenAI ties reasoning to its conversation history, on top of other safeguards. The post does not say which safeguard caused the drop.

OpenAI said on 20 November that it has hardened GPT-6.2 Astra and GPT-6.2 Sol against adversarial distillation. That means training a rival model on a frontier model's outputs without permission. In OpenAI's own test, automated attackers built on six models tried to recover reasoning from an earlier request. They succeeded 54.0% of the time on GPT-6 Astra and 0.2% on GPT-6.2 Astra. Attacks on the current request were below 1%.

Reasoning means the steps a model works through before answering. OpenAI says these traces can hold details the final answer omits. In its study, students trained on exact traces generally beat those trained on summaries. After extra reinforcement learning, the exact-trace student led only at the largest inference budget tested. In September, attackers moved encrypted reasoning between conversations and asked the model to reveal it. Now the reasoning is tied to the history that produced it, and the API drops reasoning that followed any rewritten message.

This binding sits on top of refusal training, classifiers and account investigations. OpenAI does not say which safeguard caused the drop, and the post names no outside tester. Buyers can ask who ran such tests and whether they covered tool-call arguments, which OpenAI says can leak reasoning.

A WebPulse Brief: a short report of an important event, written by the WebPulse Newsroom with AI assistance and checked against the reporting below. When there is more to explain, we follow up with a full story. How we use AI.

Reporting: OpenAI.