News

OpenAI Hardens Chain-of-Thought Defenses: Extraction Dropped from 54% to 0.2% (Nov 20, 2026)

On November 20, 2026, OpenAI published a note on adversarial distillation, disclosing new chain-of-thought protections for GPT-6.2 Astra and GPT-6.2 Sol. In an internal adaptive black-box red-team evaluation, the average attack success rate for extracting protected reasoning from earlier requests fell from 54.0% under GPT-6 Astra's original defenses to 0.2% for GPT-6.2 Astra; extraction from current requests fell from 22.9% to 0.9%. Protections span adversarial training, input and output classifiers, API-level binding of reasoning to conversation history, cross-model-family isolation, and account-level investigation.

One-line takeaway

What distillation cannot copy is not the answer but the reasoning behind it. OpenAI just turned "who may read the chain of thought, under what conditions" into a permission question at the API layer.

Background ​

On November 20, 2026, OpenAI published "Combatting adversarial distillation," describing how it detects and blocks unauthorized distillation and detailing chain-of-thought protections added to GPT-6.2 Astra and GPT-6.2 Sol.

OpenAI defines adversarial distillation as unauthorized use of a frontier model's outputs to train another model, giving the second model similar capabilities without comparable safeguards. Its example is cybersecurity: the original model may help find vulnerabilities and write defensive code while refusing to exploit them; a model trained extensively on that output, without comparable controls, may generalize those skills into offensive tasks the original model would refuse.

The framing matters because it shifts the argument from "copying outputs is fine" to "copying the reasoning path is a capability transfer."

The numbers ​

Attack typeGPT-6 Astra, original defensesGPT-6.2 Astra, new defenses
Extracting protected reasoning from earlier requests54.0%0.2%
Extracting protected reasoning from current requests22.9%0.9%
Black-box adaptive evaluation, per attack typeNot given as one figureBelow 1%

These success rates come from an internal adaptive black-box red-team evaluation aggregated over six attack models. OpenAI says extraction success against GPT-6.2 Sol was likewise below 1%, and that some protections were backported to GPT-6 Astra.

"Black-box" means the attacker sees only what an ordinary user sees. "Adaptive" means the attack strategy adjusts against the defense. That is closer to reality than a fixed attack set, but it is still not the same as an independent researcher's strongest attempt.

Why reasoning is the asset ​

OpenAI's rationale: the intermediate steps a model takes before producing an answer can contain information absent from the final answer. Those traces are more valuable for training another model than the answer itself, or than a compressed explanation.

A concrete reading: a model permitted to assist with defensive security work may show, in its reasoning, how to determine that a vulnerability is exploitable. Extracted on its own, that reasoning looks closer to method than to conclusion, and therefore transfers more easily to another model.

Four layers of defense ​

LayerWhat it does
ModelAdversarial training so the model recognizes requests that try to elicit hidden reasoning, while keeping normal explanation ability
ClassifiersInput classifiers screen for extraction intent; output classifiers scan replies and tool-call arguments for leakage
APIProtected reasoning is bound to the conversation history that produced it; if the prefix is rewritten, the reasoning after it is discarded; older model families cannot read newer models' protected reasoning
AccountDetects coordinated extraction across accounts and networks; can act on associated accounts and share indicators with other model providers

The API rule is the most consequential. Binding reasoning to its conversation history closes the path of "move the encrypted reasoning into another conversation and ask for it in plain language," an attack OpenAI says it disclosed previously. It also narrows the boundaries between users, workspaces, organizations and model families.

Note that output classifiers also inspect tool-call arguments, because sensitive content can appear in what is sent to a tool rather than in the visible answer. That matters especially for agents with tool use.

There is a practical cost too. Applications that compress or rewrite conversation history to save context now face a trade-off: keep the reasoning, or keep the freedom to rewrite the history. You cannot always have both.

How much capability transfer remains ​

OpenAI also published a cybersecurity capability-transfer study. The same pretrained student model was fine-tuned two ways: on teacher trajectories containing precise reasoning traces, and on user-visible material only (final answers, reasoning summaries, tool interactions).

Results: the precise-trace student performed better after supervised fine-tuning; after lightweight reinforcement learning it still led at the largest tested reasoning budget; but at lower reasoning budgets, the summary-trained student did better.

That is a more nuanced finding than "defenses work." It suggests traces carry something transferable, but the amount depends on training method and inference budget. It is not all-or-nothing.

For defenders, the implication is that protection lowers high-budget transfer but does not guarantee low-budget transfer is blocked. For someone seeking capability cheaply, summaries may be enough.

What the announcement changes in practice ​

First, reasoning visibility moves from product feature to permission item. Clients should ask who owns the reasoning, whether it can cross conversations, and whether a departed employee's account can still read it.

Second, batch and coordinated extraction become a compliance issue. OpenAI says it investigates account- and network-level coordination and can share indicators with other providers. Spreading extraction across hundreds of accounts does not make it invisible.

Third, teams training their own models need to check two things about data provenance: whether the source service's terms allow using its outputs for training, and whether any evasion technique was used to obtain reasoning that was not meant to be visible. If either fails, the practice is outside normal use.

Boundaries ​

  • The publication date, success rates, defenses and API behavior come from OpenAI's note and the GPT-6.2 system card; secondary coverage is broadly consistent;
  • All rates are vendor-run evaluations, not independently verified; sample sizes, evaluator design and attack budgets are not disclosed;
  • "Six attack models aggregated" and "below 1%" are official descriptions without per-item figures;
  • The capability-transfer study is also vendor-reported; its full benchmark list, student size and training recipe are not public;
  • This article contains no method for bypassing reasoning protections, evading classifiers, or conducting unauthorized distillation.

Sources ​