OpenAI says reasoning extraction success fell from 54.0% to 0.2% in its own test; the post prints no outside test
OpenAI's post of 20 November 2026 says the average success of attacks recovering protected reasoning from an earlier request fell from 54.0% on GPT-6 Astra to 0.2% on GPT-6.2 Astra, 270 times lower (derived), in a test OpenAI ran itself. The post prints no outside test.
By Parminder Kumar Sharma · LinkedIn · · 19 min read
A fall of 270 times (derived), measured by OpenAI against its own attackers
OpenAI's post, "Combatting adversarial distillation", dated 20 November 2026, prints a two row table of attack success rates. In the first row, an attack that tries to recover protected reasoning from an earlier request succeeded on average 54.0% of the time against GPT-6 Astra with its original safeguards and 0.2% of the time against GPT-6.2 Astra with its updated safeguards. That is 270 times lower and 53.8 points lower (derived: 54.0 / 0.2 = 270; 54.0 - 0.2 = 53.8). In the second row, extraction from the current request, the rate went from 22.9% to 0.9%: 25.4 times and 22.0 points lower (derived: 22.9 / 0.9 = 25.44; 22.9 - 0.9 = 22.0). This site read the page as saved at about 08:43 UTC on Saturday 21 November 2026, the day after its date.
What that does not establish. Both rows come from adaptive black box red teaming that the post introduces as an "internal adaptive evaluation using strong automated red-teaming", with the rates "aggregated across six attacker models". The post names none of the six, prints no count of attempts, does not say how a successful extraction is scored, and prints no test by anyone outside OpenAI. It does not say that real extraction attempts fell by any amount: a fall in a test is not a fall in attacks. And it compares two models, GPT-6 Astra and GPT-6.2 Astra, whose safeguards also differ, so the 270 times mixes a model change with a safeguard change.
Rounding. The page prints one decimal. A printed 0.2 could be anything from 0.15 to 0.25, which would make the first ratio anything from 216 to 360 times (derived: 54.0 / 0.25 = 216; 54.0 / 0.15 = 360). The direction is not in doubt on OpenAI's own figures; the size of the multiple is softer than 270 suggests.
Whose account this is. OpenAI sells the models it protects, and it wrote and ran this test. The post's own words are that, "Beyond intellectual property concerns", unauthorized distillation poses safety risks, so protecting OpenAI's outputs is both a safety aim and a commercial one. The post also argues for legislation and export controls. This site treats it as it would any vendor's claim about its own product. No page this site read gives an independent figure for this result. The earlier OpenAI extraction report is covered in this site's brief on the 30 September post; this one adds a defence figure to it and, as the sections below show, still no count of campaigns.
What OpenAI says it built: two levels and a binding of reasoning to its conversation
The post's threat is "adversarial distillation": an actor "may use a frontier model's outputs without authorization to train another model with similar capabilities, then deploy it without comparable protections". Its own example is hypothetical: "an adversary might distill a model's cyber capabilities and use the resulting model to help carry out cyberattacks". The post names no model that has been distilled and then used in an attack. The asset it protects is "protected reasoning", defined as "the intermediate steps a model generates internally while working toward an answer, known as its chain-of-thought". It gives the reason for the focus: "These traces can contain information the final answer leaves out, making them especially valuable training data for another model."
The defence has two levels. "At the request level, model behavior, classifier checks, and API controls limit reasoning trace extraction." "At the campaign level, we look for patterns in request-level and account-level signals that can reveal coordinated extraction attempts." The request level has three parts, per the post. Model behaviour: adversarial training so the model refuses to disclose protected reasoning, with training examples drawn from successful attacks found by "internal automated red-teaming, investigations, and responsible disclosures" from outside researchers or industry partners. Classifiers: input classifiers for extraction attempts and output classifiers for leaks "in assistant messages and tool-call arguments", applied to streaming and non streaming responses, which can block a suspected leak before it reaches the caller. API protections: these restrict how protected reasoning can be reused across interactions and models.
The API change is the most specific. For GPT-6.2 Astra and GPT-6.2 Sol, "we tie a model's protected reasoning to the conversation history that produced it", and "If an earlier message is rewritten, our API drops the reasoning that followed it". The post adds that older model families with weaker safeguards are prevented from reading the protected reasoning of the new models, and that protections were strengthened across user, workspace, organisation and model family boundaries, including closing a pathway that could expose another user's encrypted reasoning. The attacks the post's "September disclosure" describes, which this site covered in brief 204: encrypted reasoning moved from one conversation into another, then the model asked to reveal it.
What this defends and what it does not. The measured request level controls are about reasoning traces: the post says they "limit reasoning trace extraction". Distillation in general is wider. The post defines it as a model learning "from another model's answers, its reasoning, or its judgments about which answers are better and why", and the cyber study below uses the content normally visible to users as its baseline. Nothing in the post describes a control on what a user does with answers they are properly given. The stated aim is modest in its own words: "to constrain the volume and value of the data distillers are able to obtain", to raise the cost of campaigns and to shorten the response time. That is a limit on extraction, not an end to it.
The two rows, and what changed between the columns
OpenAI's table of attack success rates (adaptive black box red teaming, aggregated across six attacker models, lower is better), as printed on 20 November 2026, with this site's derived columns
| Attack, in OpenAI's words | GPT-6 Astra, original safeguards | GPT-6.2 Astra, updated safeguards | Derived |
|---|---|---|---|
| Extracting protected reasoning from an earlier request | 54.0% | 0.2% | 270 times lower, 53.8 points lower |
| Extracting protected reasoning from the current request | 22.9% | 0.9% | 25.4 times lower, 22.0 points lower |
The post's opening paragraphs carry only the first row, as an average attack success rate that "fell" from 54.0% to 0.2%. The second row sits in a table lower down and is described as "below 1% for each attack type". Both move the same way, and the second is the smaller fall, 25.4 times against 270 (derived). This is not a discrepancy, only a point about which figure a reader meets first.
Two things changed at once. The column headings read "GPT-6 Astra defender with original safeguards" and "GPT-6.2 Astra defender with updated safeguards". The model differs and the safeguards differ, and the page prints no row that separates them: no GPT-6.2 Astra with the old safeguards, and no GPT-6 Astra with the new ones. The post says it is "applying selected safeguard improvements to GPT-6 Astra as well", so a version of GPT-6 Astra with some new safeguards exists or is coming, but no figure for it is printed. The history binding described above is introduced for GPT-6.2 Astra and GPT-6.2 Sol, so it sits in the right hand column. Inference, this site's: the earlier request attack asks for reasoning produced in a previous request, and tying reasoning to the conversation that produced it is the kind of control that bears on that attack, which would fit the larger fall in the first row. The post does not say how much of any fall each control accounts for.
The ranking flips (derived). On GPT-6 Astra the earlier request attack succeeded 2.4 times as often as the current request attack (54.0 / 22.9 = 2.36). On GPT-6.2 Astra the current request attack succeeded 4.5 times as often as the earlier request one (0.9 / 0.2 = 4.5, rounded inputs). The post does not comment on this. It is a small sign that the safeguards did not fall evenly, and with rates this low and rounded to a tenth, nobody can say how stable the order is.
Only one kind of test has a number. The post describes four: static, where testers "replay a fixed set of previously successful attacks"; adaptive, where "automated AI red-teamers adjust their approach based on how the system responds"; black box, which gives testers "the same view an external user would have"; and white box, which gives them "additional information about the system to help them develop stronger attacks". Numbers are printed only for the adaptive black box test. No static result and no white box result appears. By the post's own description, the printed figure is for the attacker with less information. The post adds that the test runs "in a setup that closely mirrors production deployments of our models and safeguards"; it says nothing more about what that setup is.
GPT-6.2 Sol has no figure. The post says the defences were strengthened "in GPT-6.2 Astra and GPT-6.2 Sol", but the table names only GPT-6 Astra and GPT-6.2 Astra. Whatever was measured for Sol is not printed. This site's brief on the GPT-6.2 card covers the card's cyber evaluations, which the card says were run "with additional system-level safeguards disabled". That is a different property from how well reasoning is protected against extraction.
Three questions the evaluation design leaves open. First, "average" across six attacker models: the page does not say whether that is a mean of six rates or a pooled rate, and prints no per model rates, so one strong attacker could sit inside a low average. Second, the post says successful attacks found by internal red teaming become training examples, and that static tests replay "previously successful attacks". It does not say whether the attacks in the adaptive test were held back from training. Third, nobody outside OpenAI is described as having run or checked the test. The post does list "responsible disclosures" from outside researchers or industry partners, but as a source of training examples, not as a check on the 0.2%.
The cyber study: exact traces helped, but not at every budget
The post's second result is a study of how cyber capability transfers. "We started each training run from the same pretrained student model." The training set consisted "primarily" of trajectories from Astra level teachers on coding and defensive cybersecurity tasks, and each student was evaluated after supervised fine tuning and again after "a lightweight round of reinforcement learning in cybersecurity and coding environments". The baseline was the content normally visible to users: final responses, reasoning summaries, and tool interactions. The test was how much more transfers when the student also sees the exact reasoning traces.
The result, in the post's words: after supervised fine tuning the exact trace student "generally outperformed the student trained on reasoning summaries". After reinforcement learning it "still scored higher at the largest inference budget tested on both benchmarks shown, while the summary-trained student scored higher at lower budgets". So the advantage of exact traces after reinforcement learning is real at the largest budget and reversed at lower budgets. The post's conclusion is hedged: "protecting reasoning traces can limit cyber capability transfer, even when the underlying task is one allowed under our usage policies". It adds a line that matters for anyone who reads this as a case for hiding traces: "A second model trained on many of those benign outputs may learn skills that also apply to offensive cybersecurity tasks." That sentence is about the visible baseline. The post's claim is that exact traces add transfer, not that hiding them stops it.
What the cyber transfer study states and does not state, from OpenAI's post of 20 November 2026 as read on 21 November 2026
| Item | Stated | Not stated |
|---|---|---|
| Teacher data | "trajectories from Astra-level teacher models working on coding and defensive cybersecurity tasks", with tool interactions and final response | The number of trajectories, the tasks, which Astra level models |
| Student | "the same pretrained student model" at the start of each run | Its size, its maker, whether it is a model anyone deploys |
| Baseline | Content normally visible to users: final responses, reasoning summaries and tool interactions | How much a student gains from the baseline over an untrained one (the caption describes two runs per panel) |
| After supervised fine tuning | The exact trace student "generally outperformed" the summary student | How much, on which tasks, what "generally" excludes |
| After reinforcement learning | The exact trace student was still higher at the largest inference budget tested; the summary student was higher at lower budgets | Where the lines cross, by how much, and why |
| Benchmarks and scores | "both benchmarks shown" | The names of the two benchmarks and every score: the chart's data is not in the saved text |
| Number of runs | Two runs per panel, per the caption | Seeds, repeats, any spread |
The chart. The post shows the result as a chart. Its caption says each panel compares two runs, and the text says "both benchmarks shown". The saved copy of the page holds that chart as an empty, script rendered element, so this site has the caption (blue lines for the user visible baseline including reasoning summaries, yellow lines for exact reasoning traces, dashed for after supervised fine tuning, solid for after reinforcement learning) and no data. The page text names no benchmark for the chart and prints no score from it. This site prints none.
A footnote and a competitor. The footnote says Anthropic's related work "reported a similar advantage when training on full reasoning transcripts rather than summaries on SWE-bench Verified" and that OpenAI's result "extends to frontier cybersecurity". This site did not read that Anthropic work. OpenAI competes with Anthropic, and a rival's work is being cited as support. The post ends the section with a promise, not a finding: "We will continue studying how much additional capability transfers across different student models and distillation training recipes."
Real campaigns: a weekly scale of accounts, two names, no count
The post does give some real world detail. "We have seen multiple unauthorized adversarial distillation campaigns on our platform this year, and shared initial findings in our September report." Its enforcement "ranged in scale from thousands to tens of thousands of accounts" each week, "with the rate increasing significantly this past summer". The post does not say whether the rate is the number of accounts, the number of campaigns or the time between detection and action. It says OpenAI tracks "the prevalence of distillation abuse, as well as the time between detection and enforcement", and prints neither measure.
Two parties are named. "In February, we briefed Congress on activity indicating that DeepSeek was attempting to distill U.S. frontier models." In September, the post says, OpenAI reported disrupting a campaign to extract protected reasoning and "attributed a core part of that activity to individuals associated with Moonshot AI, the developer of Kimi". The February statement is an allegation OpenAI makes about activity it says indicated an attempt; the post gives no evidence for it and this site read none of the pages the post links. The September attribution is the one brief 204 examined: a core cluster tied to individuals associated with Moonshot AI, with no indicator published. The post closes this thread with "We plan to share more soon about recent campaigns we've disrupted." It names no campaign date, count of attempts or count of campaigns.
Against brief 204 the new post adds a scale of accounts per week, the DeepSeek reference, and a defence figure. It still gives no count of campaigns, no dates, no requests, no captured volume and no indicators. Brief 204 recorded 16,000 attempted requests from over 4,000 users on 24 and 25 July, and related prompt pattern activity across more than 15,000 users. The two sets of numbers do not compare: users against accounts, a two day spike against weekly actions, attempts against enforcement. This site derives no ratio between them. The new post is dated 51 days after the 30 September post covered there (derived: 30 September to 20 November, with 30 September taken from that brief).
What happens to the people on the other side of a false alarm is thin. The page says adversarial training works "while preserving responses to legitimate requests for explanations and assistance", says an output classifier can block a suspected leak, and says OpenAI enforces "against the network of accounts". It prints no measure of legitimate requests blocked, no false positive rate, and no route to challenge a block or an enforcement. Brief 204 raised the same question about the 15,000 users; the September post gave no answer to it.
The national security section is advocacy and is labelled as OpenAI's position: it says it supports legislation that would "give the U.S. government more tools to identify and sanction foreign actors that illicitly extract capabilities from American AI models", and that it has told Congress it supports the Chip Security Act, the AI OVERWATCH Act and the MATCH Act. It argues that stronger export controls limit hardware and that distillation defences "make it harder to acquire some of the same capabilities from models already trained". None of that is evidence about the 0.2%.
Labels doing the work of evidence
Five phrases in the post read like measurements and are not, taken alone. The table gives what each tells a reader and what it leaves out.
Labels in OpenAI's post of 20 November 2026 and the question each one skips
| Label | What it tells you | What it leaves out |
|---|---|---|
| Strong automated red-teaming | The attackers are software, and OpenAI calls them strong. | Which models, how many attempts, how strength was judged, whether they had the budget a funded campaign has. |
| Updated safeguards | Something changed between the columns. | Which change did the work, since the model changed too. |
| Adaptive | The attackers react to the system's replies. | How many rounds they were given, and whether the test attacks overlap with the training attacks. |
| Protected reasoning | A named asset with controls around it. | How much of the value sits in it: the post says exact traces add transfer, and also that visible outputs may teach offensive skills. |
| Preserving legitimate requests | A design goal for the training. | Any measurement of it. |
What the post asks of other labs
The post closes with four steps for other developers and platforms. In the post's words, in order: "Protect private reasoning across the full interaction", including leaks in tool call arguments and reuse of encrypted reasoning across conversations and models; keep "successful extraction attack logs" as regression tests, and replay them against updated defences; "Combine request-level mitigations with campaign-level investigation and enforcement"; and coordinate safeguard requirements and enforcement with deployment partners and other developers.
These are OpenAI's lessons from its own system. The post offers no test of whether they work elsewhere, and the evaluation described above is OpenAI's own. It also says OpenAI shares indicators of compromise with other model providers, shares findings through the Frontier Model Forum and government channels, and has a "consistent-standards policy" that sets shared standards for "providers offering our models", which this site reads (inference) as a commercial term as well as a safety one. The post names no provider that has adopted any of this.
What the post states, and what it does not
OpenAI's post of 20 November 2026 read against the questions a risk owner asks
| Question | Stated | Not stated |
|---|---|---|
| Did extraction attacks fall? | In OpenAI's own adaptive black box test: 54.0% to 0.2% (earlier request) and 22.9% to 0.9% (current request), across six attacker models. | The attackers' names, attempts, scoring, any white box or static result, any outside test. |
| Is it the safeguards? | Updated safeguards on GPT-6.2 Astra; selected improvements now going to GPT-6 Astra. | A row separating model change from safeguard change. |
| Does it hold for Sol? | Defences strengthened in GPT-6.2 Astra and GPT-6.2 Sol. | Any Sol figure. |
| Is it happening? | Multiple campaigns this year; weekly enforcement from thousands to tens of thousands of accounts. | A count of campaigns, dates, requests, data captured, indicators. |
| By whom? | DeepSeek (Congress briefed in February) and Moonshot AI associates (September), as OpenAI says. | Evidence for either, or any other actor. |
| Does hiding traces stop transfer? | Protecting traces "can limit" transfer; exact traces added capability at the largest budget. | The size of the effect, the benchmarks, the scores, and any control with no distillation. |
| What does it cost legitimate users? | Training aims at "preserving responses to legitimate requests". | A measure of that, or an appeal route. |
What to do, in order
These are this site's suggestions, except where marked as the post's own advice. The post is a vendor's account of its own defences, and nothing here depends on its number being right.
Take this with you
In the order worth doing
- If you run a model API, separate request level controls from campaign level detection, as the post describes (the post's own advice), and keep each successful extraction as a regression test (also the post's advice).
- Treat exact reasoning traces as sensitive output. Decide what your API returns, summaries or traces, and check tool call arguments and streamed output for leaks, as the post says its own classifiers do.
- If you buy model access, ask the vendor for the design behind any attack success rate: attacker models, attempts, how success is scored, black box or white box, whether the attacks were held back from training, and who ran it.
- For defenders, treat a distilled copy of a capable model with its safeguards stripped as a plausible attacker tool class. That is the post's own hypothetical, and it names no such model in use, so plan for faster assisted attack work without assuming it is already here.
- Read the model card for what is protected, and check which of your own data flows return reasoning at all. This site's brief on the GPT-6.2 card is a starting point for what OpenAI says it measured.
The question the post leaves open
The post says attacks that succeed in red teaming and investigations become training examples, and that the 0.2% comes from an adaptive test run by OpenAI's own attackers. If the attacks that worked are logged and fed back into the model, how much of the 270 times is a defence generalising to attacks it has not seen, and who outside OpenAI has checked?


