sysnote
News

OpenAI says GPT-6.2 drops edited reasoning to stop distillation; extraction success fell from 54% to 0.2%

On November 17 we wrote that the Responses API drops reasoning items for gpt-6.2-astra and gpt-6.2-sol when an application edits earlier conversation history, moves the conversation to certain other models (including a downgrade), or changes the end user or organization. The GPT-6.2 guide describes the rules, and the prefix guide says only that the API drops the items "as part of our safeguards." OpenAI's November 20 report, Combatting adversarial distillation, gives the reason. The rules are meant to stop people from extracting the model's hidden reasoning and training another model on it.

OOpenAI@OpenAI

Distillation can reproduce powerful AI capabilities without the safeguards built around them.

We’re sharing new research on our work to combat distillation of our models, how these capabilities transfer, and how we detect and disrupt coordinated attempts.

These findings are…

November 20, 2026

OpenAI calls that hidden chain of thought protected reasoning. It moves between API requests in encrypted form, so an application can hand it back on the next turn without seeing it. In a September 30 disclosure OpenAI described operators who copied encrypted reasoning from one conversation into another and asked the model there to decrypt and transcribe it. On July 24 and 25 it counted 16,000 requests with that extraction pattern from more than 4,000 users, which it says were attempts, not necessarily successful ones. It attributed a core cluster of the activity to people associated with Moonshot AI, the developer of Kimi.

Each API rule closes a route

The new report's section on API-level protections lines up with the migration guide's rules:

  • Prefix edits. "With GPT-6.2 Astra and GPT-6.2 Sol, we tie a model's protected reasoning to the conversation history that produced it. … If an earlier message is rewritten, our API drops the reasoning that followed it." Reasoning copied into another conversation, as in the September campaign, no longer has the history it was bound to.
  • Model changes. "We also prevent older model families with weaker safeguards from reading GPT-6.2 Astra and Sol's protected reasoning." The September disclosure says outside researchers had also reported cross-model attack paths.
  • Users and organizations. OpenAI says it strengthened protections "across user, workspace, organization, and model-family boundaries," including closing a pathway that could expose another user's encrypted reasoning.

Two checks sit outside the request format. Output classifiers look for leaked reasoning in assistant messages and in tool-call arguments, because a model can put its reasoning into a tool call without writing it in the answer. The GPT-6.2 system card adds "checks that can hold streamed responses if they might expose private model reasoning." Neither document says how often a legitimate stream is held, or for how long.

OpenAI's numbers

In OpenAI's adaptive black-box red-teaming, automated attackers built on six attacker models adjusted their attempts to the system's responses. The table gives the average attack success rate:

Attack GPT-6 Astra, original safeguards GPT-6.2 Astra, updated safeguards
Extract reasoning from an earlier request 54.0% 0.2%
Extract reasoning from the current request 22.9% 0.9%

The system card puts both GPT-6.2 Astra and Sol below 1% for both attack types. These are OpenAI's own evaluations. The report also says weekly enforcement actions against distillation campaigns have ranged from thousands to tens of thousands of accounts, with the rate rising significantly this summer.

The transfer study is less clear-cut than Anthropic's

To show why exact traces are worth protecting, OpenAI fine-tuned the same pretrained student model on trajectories from Astra-level teachers working on coding and defensive security tasks. One student saw the teacher's exact reasoning. The baseline saw what a user normally sees: final responses, reasoning summaries and tool interactions. After supervised fine-tuning, the exact-trace student "generally outperformed" the summary student. After a light round of reinforcement learning, it scored higher only at the largest inference budget tested on both benchmarks shown, and the summary-trained student scored higher at lower budgets. The report's text gives no numbers for the study and doesn't name the student model or the dataset size.

Anthropic's October 12 post gave a single, cleaner gap: a pretrained Claude model trained on Claude Opus 4.8 transcripts reached 80.4% on SWE-bench Verified with the full reasoning and 71.8% with summarized reasoning. OpenAI's footnote cites that result and says its own study shows the advantage "extends to frontier cybersecurity." That holds for OpenAI's fine-tuned student. After reinforcement learning, its own description limits the advantage to the highest budget it tested.

Both providers now penalize rewritten history

Anthropic's version is preserved thinking, which we covered with Claude Fable 5.2. A Claude thinking block is bound to the system prompt, tools and earlier messages; on Fable 5.2 a mismatch returns a 400 error unless the developer opts into dropping the blocks. Older Claude models can't read the newer models' reasoning either. OpenAI drops the affected reasoning items without failing the request and lists them in input_transformations.

The result is similar at both providers. An agent harness that shortens old tool output, re-renders its system prompt, swaps the tool list or runs its own compaction now loses earlier reasoning, either as an error on Fable 5.2 or as repeated reasoning and extra tokens on GPT-6.2. OpenAI's prefix guide lists the append-only replacements: a configuration_update item instead of a new reasoning.effort or text.format, additional_tools items or allowed_tools instead of a changed tools list, a new developer message at the end instead of an edited first one. Its own compaction through /responses/compact or context_management keeps the reasoning that follows.

One question is left for integrations still on the previous model. The report says OpenAI is "now applying selected safeguard improvements to GPT-6 Astra as well," without saying which. If that includes the prefix rule, GPT-6 Astra agents that rewrite history will start losing reasoning too. The API changelog has no entry about it as of November 21.

Comments

No comments yet. Be the first.

Write a comment

Comments are published after review.