Skip to content
AI EmployeeSuper-AgentsAgent-to-AgentBenchmarksPricingBlog

OpenAI's Distillation Defenses: 54% to 0.2%

At a glanceQuick answers
What is the headline number?
Reasoning extraction from an earlier request fell from 54.0% on GPT-6 Astra to 0.2% on GPT-6.2 Astra, in OpenAI’s own test.
Which models get the new safeguards?
GPT-6.2 Astra and GPT-6.2 Sol, with selected improvements also applied to GPT-6 Astra.
Who checked the figures?
No one outside OpenAI yet. They come from its internal red-teaming.
Data illustration on a near-white ground: a teal vault door holding a stack of thin sheets, an orange claw stopped by a teal gate, and the figures 54.0% falling to 0.2%, labelled reasoning extraction success.
Fig 0OpenAI's own red-team figure for extracting earlier reasoning. Made by CellCog's image agent, running GPT Image 2.6.

OpenAI says it has nearly closed the door on attacks that copy its models’ hidden reasoning. In research announced on X at 19:33 UTC on November 20, 2026, the company said: “Distillation can reproduce powerful AI capabilities without the safeguards built around them.” Its post reports that adaptive attacks extracting protected reasoning from an earlier request now succeed 0.2% of the time on GPT-6.2 Astra, against 54.0% on GPT-6 Astra with its original safeguards.

We did not open openai.com for this record. The figures below come from OpenAI’s X post and two named reports of the post, by Unite.ai and MLQ.ai, which agree on every number we use.

Key points5 · 4 min full read
  1. A teal shield with a downward arrow.
    OpenAI said on November 20, 2026 that adaptive attacks extracting protected reasoning from an earlier request succeed 0.2% of the time on GPT-6.2 Astra, down from 54.0% on GPT-6 Astra.
  2. A magnifying glass over a bar chart.
    For the current request, the same test fell from 22.9% to 0.9%. Both figures come from OpenAI’s own red-teaming across six attacker models.
  3. A teacher sheet passing a copy to a student sheet.
    In an internal study, a student model trained on exact reasoning traces generally beat one trained on summaries on cyber tasks, which is why OpenAI guards the traces.
  4. Two chat bubbles joined by a padlock.
    In the API, GPT-6.2 Astra and Sol tie protected reasoning to the conversation that produced it, and older model families cannot read it.
  5. A group of user icons stopped by a sign.
    OpenAI says weekly enforcement reaches thousands to tens of thousands of accounts, and names DeepSeek and individuals associated with Moonshot AI in earlier findings.

§ 01The numbers OpenAI reports

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%
Table 1Attack success in OpenAI’s adaptive black-box red-teaming, across six attacker models (lower is better)
Reasoning-extraction success by attack and model, percent (OpenAI's own test)Bar chart of OpenAI's reported extraction success: earlier request on GPT-6 Astra 54.0, on GPT-6.2 Astra 0.2 highlighted, current request on GPT-6 Astra 22.9, on GPT-6.2 Astra 0.9Earlier, GPT-6 Astra54.0Earlier, GPT-6.2 Astra0.2Current, GPT-6 Astra22.9Current, GPT-6.2 Astra0.9Reasoning-extraction success by attack and model, percent (OpenAI's own test)Bar chart of OpenAI's reported extraction success: earlier request on GPT-6 Astra 54.0, on GPT-6.2 Astra 0.2 highlighted, current request on GPT-6 Astra 22.9, on GPT-6.2 Astra 0.9Earlier, GPT-6 Astra54.0Earlier, GPT-6.2 Astra0.2Current, GPT-6 Astra22.9Current, GPT-6.2 Astra0.9
Fig 1Reasoning-extraction success by attack and model, percent (OpenAI's own test)

The red-teamers were automated and adaptive: they changed approach as the system responded and saw only what an outside user would see, according to both reports.

Animation: an isometric platform with an orange jointed arm labelled Extraction attack reaching toward a teal vault labelled Hidden reasoning; the vault door closes over a stack of sheets; a tall grey bar labelled GPT-6 Astra reads 54.0% and a flat bar labelled GPT-6.2 Astra reads 0.2%; then the line Extraction fell to 0.2%, by OpenAI's own test.
Fig 2In OpenAI's own test the vault holds and extraction falls from 54.0% to 0.2%

§ 02Why the traces matter

OpenAI trained the same pretrained student model two ways on Astra-level teacher trajectories from coding and defensive security tasks: once on exact reasoning traces, once on what users see (final answers, reasoning summaries and tool interactions). After supervised fine-tuning, the exact-trace student generally did better. After a light round of reinforcement learning, it still scored higher at the largest inference budget on ExploitBench and SEC-bench Pro, while the summary-trained student did better at lower budgets.

That is the case for hiding the traces: they carry skills the final answer does not, and those skills can transfer.

§ 03What changed

Level Defense
Model Trained to recognize extraction attempts and refuse to reveal protected reasoning
Request Input classifiers for extraction attempts; output classifiers on messages and tool-call arguments
API Reasoning tied to its conversation; rewriting an earlier message drops what followed
API Older model families cannot read GPT-6.2 reasoning; a path to another user’s encrypted reasoning was closed
Campaign Account-network investigations, enforcement, and indicators shared with other providers
Table 2OpenAI’s distillation defenses, as reported on November 20, 2026

OpenAI says weekly enforcement against distillation campaigns has ranged from thousands to tens of thousands of accounts this year, rising sharply over the summer. It names DeepSeek, which it briefed Congress about in February, and individuals associated with Moonshot AI, the subject of its earlier report. It says it will share more about recent campaigns soon.

§ 04What the numbers leave out

  • Self-reported. Every figure comes from OpenAI’s internal testing. No outside group has reproduced it yet.
  • Benchmark scores. The reports give the direction of the study results, not the scores themselves.
  • Older models. GPT-6 Astra gets only selected improvements, so its exposure is not described by the 0.2% figure.

Anthropic published its own version in October, saying its changes cut what reasoning distillers extract by 10 to 100 times (our record). Two labs have now put numbers on the same fight; both sets are their own.

§ 05Sources

Frequently asked4 questions

Q1What is adversarial distillation?

Using a frontier model’s outputs without permission to train another model. OpenAI’s concern is that the copy can inherit capabilities without the safeguards of the original.

Q2Why does hidden reasoning matter so much?

OpenAI’s study found that exact reasoning traces taught a student model more than the summaries users see, on coding and defensive security tasks. The traces carry steps the final answer leaves out.

Q3Does this change anything for developers?

On GPT-6.2 Astra and Sol, rewriting an earlier message drops the protected reasoning that followed it, and that reasoning cannot be passed to older model families.

Q4Does this affect CellCog?

Not directly. CellCog’s Core lanes run on Anthropic’s Claude Opus 5.5 and its Max lanes on Fable 5.2; none route to OpenAI models.

Published 20 November 2026 All Category basics →