Article

When the Chain of Thought Becomes an Asset: Four Layers of Distillation Defense, and How to Read 54% to 0.2%

OpenAI has moved its adversarial-distillation defense onto protected reasoning, reporting extraction success falling from 54.0% to 0.2%. This analysis separates three forms of distillation, explains why traces are worth more than answers, maps what each of four defense layers blocks, and sets out the conditions under which a vendor's self-reported figure can be treated as evidence. It also examines the tension between protecting reasoning and keeping it auditable.

One-line takeaway

You can copy an answer; you cannot copy a method unless the method is handed over. OpenAI's move turns the reasoning process from published content into conversation-private state.

Three forms of distillation, three different problems ​

Before asking whether distillation can be stopped, separate what is being distilled.

FormTraining materialWhat transfersDifficulty
Answer distillationFinal responsesConclusions, style, domain factsLow: outputs are public
Summary distillationReasoning summaries, explanationsA coarse version of the problem-solving frameMedium
Trace distillationFull chains of thoughtMethods, judgment paths, failure modesHigh: this is the current focus

The first two teach a model how the teacher speaks. The third teaches how it thinks. For most tasks the first two suffice. For tasks that require multi-step judgment — vulnerability analysis, complex code changes, scientific reasoning — the third is where the capability actually lives.

OpenAI singles out the third because a trace may contain information the final answer omits. A model might write "not recommended" without mentioning that it tried three paths and two tripped a check. That omitted part is the most trainable material.

Why traces are worth more ​

An answer is a worked solution; a trace is the scratch paper. The scratch paper carries dead ends, crossed-out assumptions, temporary arithmetic. From the solution you learn "write this step this way." From the scratch paper you learn "why this way, and what happens otherwise."

That wider coverage is what makes traces valuable for training another model, and it explains OpenAI's capability-transfer result: students fine-tuned on precise traces generally outperform those trained only on visible material.

Reading 54% to 0.2% ​

The contrast is striking, but it has to be read inside its conditions.

It is an adaptive black-box evaluation. Black-box means the attacker sees only what a user sees; adaptive means the attack adjusts to the defense. Closer to reality than a fixed attack set, still not an independent researcher's best attempt.

It is aggregated over six attack models. 0.2% is an average, not a ceiling every attack reached. The distribution is not published.

Two rates cover two paths. Extraction from earlier requests fell 54.0% to 0.2%; from current requests, 22.9% to 0.9%. The larger drop on the first path suggests the defense bites hardest on cross-turn transport of reasoning.

It is vendor-reported. Sample sizes, evaluator design and attack budgets are not disclosed, and there is no alignment with external evaluation. Reading 0.2% as "trace extraction is solved" overstates it. Reading it as "the cost of this particular path rose sharply" is fair.

What each layer blocks ​

LayerBlocksDoes not block
Model (adversarial training)Direct requests to recite hidden reasoningMulti-turn requests that approach it gradually
Classifiers (input and output)Obvious extraction intent; leakage in replies; leakage in tool argumentsSemantically transformed requests, and formats the classifiers do not cover
API (session binding)Moving encrypted reasoning across sessions; older families reading newer reasoningSide channels during legitimate use within one session
Account (coordinated investigation)Large-scale, multi-account extractionLow-frequency, slow extraction from a single account

Stacked, the four layers turn a one-shot extraction into an engineering problem that requires long horizons, distribution and continuous adaptation against classifiers. That raises cost; it is not an absolute block.

The API design deserves separate mention: protected reasoning is bound to the history that produced it, and rewriting the prefix discards what follows. This solves two problems at once. It prevents reasoning from being carried elsewhere and recited, and it defeats "edit the history, then let it continue thinking." The side effect is engineering cost: any application that rewrites history must accept discarded reasoning and a model that re-thinks.

How much transfer actually falls ​

OpenAI's transfer study traces a curve, not a point:

  • After supervised fine-tuning, precise-trace students did better on both reported benchmarks;
  • After lightweight reinforcement learning, they still led at the largest tested reasoning budget;
  • At lower reasoning budgets, summary-trained students overtook them.

The reversal is the interesting part. Traces raise the ceiling more than the floor. Students learn longer thinking paths: beneficial when budget is ample, costly when it is tight.

For defenders, protection lowers high-budget transfer but gives no guarantee about low-budget transfer. For anyone seeking capability cheaply, summaries may be enough.

Three implications for the industry ​

Reasoning visibility is contracting in closed APIs. Showing the thinking process used to be a selling point. It is now an asset that needs boundaries by session, organization and model family. That changes downstream tooling — for example, pipelines that parse chain-of-thought as a structured field.

Open-source distillation practices face clearer compliance pressure. Training a small model on a frontier model's outputs is not automatically a violation. The combination of "unauthorized outputs plus evasion of protections" is being explicitly ruled out. The open-source community needs its own norm here.

Evaluation transparency becomes contested. When a vendor announces a 97-point drop without methodological detail, outsiders cannot tell whether the defense is strong or the evaluation is gentle. More such numbers mean more demand for independent evaluation.

A tension worth naming ​

There is a quieter contradiction: the chain of thought is both an asset to protect and a handle for safety monitoring.

A common practice is to let an external monitor read the model's reasoning to judge whether it is trying to circumvent restrictions. If reasoning is strictly protected and unreadable, that channel narrows. OpenAI's own system card notes that GPT-6.2 Astra was in some tests better at evading chain-of-thought monitoring.

Protecting reasoning from distillation and keeping reasoning auditable pull in opposite directions. What is needed is graded permission: hidden from ordinary callers, open to internal monitoring, recorded for audit. No public standard for that grading exists; each vendor designs its own.

A practical corollary: if visibility depends on session binding, then "export a conversation and continue elsewhere" becomes fragile. Teams building multi-turn applications should treat "can the history be rewritten" as a design constraint from the start, not a discovery after launch.

Boundaries ​

  • OpenAI's note, success rates, defense layers, API behavior and transfer findings come from the official note and the GPT-6.2 system card;
  • All figures are internal vendor evaluations, not independently verified, and are not presented here as industry benchmarks;
  • "Six models aggregated" and "below 1%" are official descriptions without per-item figures;
  • Benchmark names, student model size and training recipe in the transfer study are not fully public; only disclosed parts are cited;
  • The system card's remark on evading chain-of-thought monitoring is limited to specific test conditions;
  • This article contains no method for bypassing reasoning protections, evading classifiers, or conducting unauthorized distillation.

Sources ​