SegurançaImportante

OpenAI details how it guards reasoning traces against distillation, and what that means for the API

OpenAI says reasoning-extraction attacks on GPT-6.2 Astra succeed under 1% of the time in its own tests, and explains why edited conversations lose their reasoning.

Promptea Editorial4 min de leitura

OpenAI on Friday published a detailed account of how it tries to stop other developers from training models on its outputs without permission, a practice it calls adversarial distillation. The post, Combatting adversarial distillation, focuses on one asset in particular: the hidden chain-of-thought of its reasoning models, which OpenAI calls protected reasoning. It also explains several API behaviors that developers building on GPT-6.2 Astra and Sol will run into.

What OpenAI is protecting, and why

Distillation itself is a standard technique: a smaller or newer "student" model learns from a stronger "teacher". OpenAI's objection is to doing it without authorization against a frontier model, because the student can inherit capabilities, including dual-use ones in cybersecurity or biology, without the safeguards the original model was shipped with.

The post argues that raw reasoning traces are worth more to a would-be copier than the summaries OpenAI shows users. In an internal study, OpenAI trained students on trajectories from an Astra-level teacher for coding and defensive security tasks. Students trained on exact reasoning traces beat those trained on summaries after supervised fine-tuning. After reinforcement learning the picture was more mixed: the exact-trace student still scored higher at the largest inference budget on both benchmarks OpenAI used (ExploitBench and SEC-bench Pro), while the summary-trained student did better at lower budgets. These are OpenAI's own experiments; they have not been independently reproduced.

The numbers OpenAI reports

OpenAI says it combines adversarial training (teaching the model to decline to reveal its protected reasoning while still explaining its answers), input and output classifiers that cover both messages and tool-call arguments, and investigations of coordinated account networks. In its own adaptive black-box red-teaming, it reports:

  • Extracting reasoning from an earlier request: average attack success rate fell from 54.0% against GPT-6 Astra with its original safeguards to 0.2% against GPT-6.2 Astra with updated safeguards.
  • Extracting reasoning from the current request: from 22.9% to 0.9%.
  • GPT-6.2 Astra stayed below 1% for every attack type tested, according to the company. OpenAI says it is applying selected improvements back to GPT-6 Astra.
  • Weekly enforcement actions against extraction campaigns have ranged from thousands to tens of thousands of accounts, with the rate rising significantly over the summer.

All of these figures come from OpenAI's internal evaluations. The company designed the attacks and the scoring, so they show the direction of travel rather than an independently verified level of protection.

What changes for API developers

The practical part of the post explains rules that arrived with GPT-6.2. OpenAI's API changelog already noted on November 17 that GPT-6.2 changes how the Responses API preserves reasoning when a conversation prefix is edited. Friday's post gives the reason and the specifics:

  • With GPT-6.2 Astra and 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.
  • Older model families with weaker safeguards cannot read GPT-6.2 Astra and Sol's protected reasoning.
  • Protections now apply across user, workspace, organization and model-family boundaries, and OpenAI says it closed a pathway that could have exposed another user's encrypted reasoning.

For most applications this is invisible. It matters for agent frameworks and evaluation harnesses that edit history in place, for example trimming or rewriting an earlier turn to save context, retrying from a modified prefix, or handing encrypted reasoning from a GPT-6.2 run to an older model in a mixed-model pipeline. After such an edit, the model continues without the reasoning it had built up, which can change behavior and should be expected rather than debugged as a regression. If an agent depends on carried-over reasoning, append corrections as new turns rather than rewriting old ones, and do not plan on passing GPT-6.2 reasoning items to older model families.

The post does not announce new identity-verification requirements, rate limits or pricing changes.

The policy context

OpenAI recaps two earlier episodes. In September 2026 it disclosed disrupting a campaign to extract protected reasoning and attributed a core part of it to individuals associated with Moonshot AI, the developer of Kimi. In February 2026 it briefed Congress on activity it said suggested DeepSeek was trying to distill US frontier models. Both attributions are OpenAI's; Promptea could not find a public response from either company in today's material.

The company also uses the post to lobby. It says it supports legislation giving the US government more tools to sanction foreign actors who illicitly extract model capabilities, names the Chip Security Act, the AI OVERWATCH Act and the MATCH Act, and frames distillation defenses as a complement to export controls. It describes a "consistent-standards" policy that sets shared safeguard and enforcement expectations for providers that offer its models, and says it shares indicators of compromise with other labs, including through the Frontier Model Forum.

There is a tension here worth naming. Hiding raw reasoning is a defensible security measure, but it also means developers and researchers see less of how these models reach their answers. OpenAI's own data shows why it values the traces, and that same data is why outside users may want them.

Bottom line

This is not a new product. It is OpenAI explaining, with numbers, how tightly it now guards reasoning traces and how that shapes the API. Developers on GPT-6.2 should treat edits to conversation history as resetting the model's reasoning, and should read the vendor-reported attack-success figures as claims awaiting outside testing.

Por que isso importa

  • Agents that rewrite earlier turns on GPT-6.2 Astra or Sol lose the reasoning that followed, which can change behavior mid-task.
  • Mixed-model pipelines cannot hand GPT-6.2 protected reasoning to older OpenAI model families.
  • The post signals that frontier labs are treating reasoning traces as a security asset, tying distillation to export-control policy.

Pontos principais

  • OpenAI reports internal attack success for extracting earlier-request reasoning fell from 54.0% (GPT-6 Astra) to 0.2% (GPT-6.2 Astra).
  • Weekly enforcement has hit thousands to tens of thousands of accounts, OpenAI says.
  • Rewriting an earlier message drops the subsequent reasoning in the API.
  • All figures are vendor-reported; no independent evaluation is available yet.
Tags:
  • distillation
  • reasoning models
  • api
  • ai security
  • policy
Empresas:
  • OpenAI
  • Moonshot AI
  • DeepSeek
Modelos:
  • GPT-6.2 Astra
  • GPT-6.2 Sol
  • GPT-6 Astra
OpenAI on adversarial distillation and GPT-6.2 reasoning rules · Promptea