# OpenAI's anti-distillation results need a threat model, not a victory lap

OpenAI says earlier-turn extraction fell from 54.0% on GPT-6 Astra to 0.2% on GPT-6.2 Astra in one adaptive test.[1] The result covers a specific attack surface and remains vendor-reported.

OpenAI disclosed new defenses against adversarial distillation on November 20, 2026. The stack combines model training, input and output classifiers, conversation-bound reasoning, campaign detection and account enforcement.[1][2]

In OpenAI's adaptive black-box test, earlier-turn reasoning extraction fell from 54.0% on GPT-6 Astra with its original safeguards to 0.2% on GPT-6.2 Astra with the updated stack. Same-turn extraction fell from 22.9% to 0.9%.[1] These are internal results across six attacker models, not an independent guarantee against every extraction method.

The practical lesson extends beyond one provider. Protect reasoning artifacts across messages, tool arguments and model boundaries; monitor coordinated traffic; retain successful attacks as regression tests; and describe every result with its attacker budget, interface and success criterion.[1][4]

## What changed in GPT-6.2

OpenAI's November 20 disclosure describes a layered defense for GPT-6.2 Astra and Sol. Adversarial training teaches the models to refuse requests for protected reasoning while preserving ordinary explanations. Input classifiers look for extraction attempts, and output classifiers inspect assistant messages and tool-call arguments for leaks.[1]

The API now ties protected reasoning to the conversation history that produced it. If a prior message changes, the API drops later reasoning that depended on the old prefix. OpenAI also says older model families with weaker safeguards cannot read GPT-6.2 reasoning artifacts.[1]

That boundary addresses a concrete failure mode. In September, OpenAI reported a campaign that copied encrypted reasoning between conversations and asked the model in another conversation to decrypt and transcribe it in visible text. The operators did not break encryption or access stored conversations, but OpenAI linked a core cluster to individuals associated with Moonshot AI and said the broader activity involved more than 15,000 users.[3]

Request controls are paired with campaign detection. OpenAI joins request and account signals, investigates coordinated activity, enforces across related accounts and shares indicators with model providers. It says weekly actions have ranged from thousands to tens of thousands of accounts.[1] Those figures describe enforcement scale, not the number of successful extractions.

## Why exact reasoning traces matter

Distillation trains a student model on outputs from a teacher. Legitimate teams use it to compress models or transfer behavior. An unauthorized campaign can use the same method to reproduce capabilities while leaving the teacher's policy and deployment safeguards behind.[1][4]

Exact reasoning traces are attractive training data because they expose intermediate decisions that a final answer or short summary omits. OpenAI compared students trained from the same starting model on trajectories containing reasoning summaries with students trained on exact traces instead. Its cyber study found an advantage for exact traces after supervised fine-tuning and at the largest tested inference budget after reinforcement learning, while the summary-trained student scored higher at lower post-RL budgets.[1]

The study supports a narrow claim: protecting exact traces can reduce one channel of cyber-capability transfer. It does not show that final answers, tool trajectories or reasoning summaries are harmless training data. A large campaign can still collect useful supervision from ordinary outputs.[1]

Model operators should therefore avoid treating hidden reasoning as the only asset. Training examples, tool outputs, graders, preference labels and high-value answer distributions can all carry behavior worth copying. The inventory should cover every output channel the service exposes, including streams, logs, batch exports and partner-hosted APIs.

## What the 0.2% result does and does not prove

OpenAI reports that six model-based attackers adaptively tested two extraction paths in a black-box setting. Against GPT-6 Astra with its original safeguards, earlier-turn extraction succeeded 54.0% of the time and same-turn extraction 22.9%. Against GPT-6.2 Astra with the updated stack, the rates were 0.2% and 0.9%.[1][2]

This is stronger evidence than replaying a fixed attack set because the attackers adjusted to the defense. It also tests a production-like stack rather than model behavior alone. The disclosed comparison gives teams concrete regression targets for state replay and same-response leakage.[1]

The numbers remain an internal OpenAI evaluation. The article does not publish the attack prompts, sample counts per condition, confidence intervals, attacker-model identities or a reproducible harness. Readers cannot independently calculate how sensitive the result is to the attack budget or interface details.[1][2]

The comparison also changes both model generation and safeguards: GPT-6 Astra with original controls versus GPT-6.2 Astra with updated controls. It does not isolate which layer produced the gain or show how GPT-6 Astra performs after the selected improvements OpenAI says it is applying.[1]

Independent research explains why those details matter. A 2026 benchmark found that extraction attack rankings depend on the metric and that defenses vary across attacks. It argues for matched models, data, budgets and evaluation conditions, while measuring both the defender's security goal and the resulting student's capability and fidelity.[4]

A buyer should read 0.2% as evidence that this defined attack became much harder in OpenAI's test. It should not be presented as a 99.8% guarantee against model extraction, a proof that no reasoning leaks remain, or an independent validation of the full service.

## An operating plan for model API providers

- Map exposed assets and channels. Include visible answers, reasoning summaries, encrypted state, tool arguments, streams, logs, batch files, SDK caches and partner-hosted endpoints.
- Write separate threat models for trace recovery, capability imitation, provenance removal and coordinated account abuse. Specify the attacker's query budget, data budget, interface access and local processing.[4]
- Build one end-to-end regression for each known extraction path. Preserve successful red-team, production and responsible-disclosure cases, then replay them after changes to models, classifiers, APIs or streaming behavior.[1]
- Inspect output before release. Cover assistant text and tool-call arguments, and decide whether a suspected stream must be buffered or held rather than leaked token by token.[1][3]
- Bind private state to the user, organization, model family, conversation and immutable history it belongs to. Reject cross-boundary replay and discard dependent state when history changes.[1][3]
- Aggregate signals over time. Rate limits alone miss distributed campaigns, while a single unusual request proves little. Join patterns across accounts and infrastructure under documented privacy and retention rules.
- Evaluate adaptive attackers after every major mitigation. Report operational costs such as false positives, latency and blocked legitimate work. For extraction defenses, pair the security metric with the resulting student's capability and fidelity.[4]
- Verify deployment parity. First-party, cloud-partner and reseller endpoints need equivalent state binding, classifier coverage, streaming controls and enforcement paths.[1][3]
- Prepare coordinated response. Define when to restrict accounts, preserve evidence, notify partners and share indicators without publishing details that improve the attack.
- Commission independent testing for material claims. Give reviewers enough access to reproduce the stated threat model and record gaps where disclosure would create new risk.

## Questions buyers should ask providers

- Which extraction objective did you test: recovering exact traces, matching behavior, transferring benchmark capability or evading provenance detection?
- What could the attacker see and change? Ask about streaming, prefilling, tool arguments, conversation replay, encrypted state and cross-model access.
- How many queries, accounts and distinct prompts did the attacker receive, and could it coordinate across them?
- Were attacker models allowed to adapt to failed attempts, rewrite outputs or train a student before the defense was scored?[4]
- Does the result compare the same model with and without the defense, or different model generations and safeguard stacks?
- What benign requests were blocked, delayed or degraded, especially legitimate debugging and requests for explanations?
- Do cloud and regional deployments receive the same protections and enforcement response as the first-party API?[1][3]
- Which parts were independently reproduced, and which evidence remains vendor-reported?

## Limits and follow-up watches

OpenAI has shown a large reduction on two adaptive reasoning-extraction tests, but independent reproduction is unavailable. Watch for a fuller evaluation protocol, external red-team results and the promised disclosure about additional campaigns.[1]

Watch whether the selected safeguards reach GPT-6 Astra and partner deployments, and whether OpenAI reports the updated baseline. A before-and-after result on the same model would help separate model-generation effects from safeguard changes.[1]

Research is moving toward lifecycle evaluations that connect acquisition, defenses, adaptive rewriting and student training. Providers should expect buyers and auditors to ask whether a defense merely hides a signal or actually reduces transferred capability under a stated budget.[4]

## Frequently asked questions

- Is all distillation malicious? No. Distillation is a standard technique for training a smaller student from a teacher. The security issue is unauthorized extraction or capability transfer that bypasses agreed safeguards.[1][4]
- Did attackers decrypt OpenAI's stored reasoning? OpenAI says the September campaign did not break encryption or compromise stored conversations. It manipulated model interactions to reproduce reasoning in visible form.[3]
- Does the 0.2% result mean GPT-6.2 cannot be copied? No. It measures earlier-turn protected-reasoning extraction in one internal adaptive black-box evaluation.[1]
- Why inspect tool arguments? A model can place private reasoning in an argument even when its final answer contains no leak.[1]
- What should a small API provider do first? Inventory output channels, bind private state to its original context, retain known attacks as tests and monitor patterns across requests rather than relying only on per-call filters.
- What evidence would improve confidence? A documented query and data budget, interface profile, student capability and fidelity results, plus independent reproduction, would make the claim easier to evaluate.[4]

## Sources

- [[1] OpenAI: Combatting adversarial distillation](https://openai.com/index/combatting-adversarial-distillation/)
- [[2] OpenAI: GPT-6.2 Astra and Sol system card](https://deploymentsafety.openai.com/gpt-6-2-astra-sol)
- [[3] OpenAI: Disrupting a coordinated model-distillation campaign](https://openai.com/index/disrupting-a-coordinated-model-distillation-campaign/)
- [[4] Liu et al.: Lifecycle benchmark of black-box model extraction](https://arxiv.org/html/2610.00839)
