OpenAI Distillation Defenses Cut Reasoning Theft to 0.2%, With a Catch for API Users

OpenAI says attacks extracting protected reasoning fell from 54.0% to 0.2% on GPT-6.2 Astra, a roughly 270-fold drop. We put the figures in context, compare them with our October coverage and explain why edited conversation history now costs developers reasoning, tokens and latency.

Christopher Ort
By i10X News Desk · Reviewed by Christopher Ort · 3 min read
Share
A padlocked steel mesh cage door in front of blurred rows of server racks in a data center.
Image: AI-generated, i10X
On this page
  1. Why reasoning traces are the prize
  2. The numbers in context: our calculation
  3. From reports to OpenAI's own account
  4. What changes for developers using the API
  5. What is still open

OpenAI says it has made the reasoning of its newest models far harder to steal. In a post on combatting adversarial distillation published on 20 November, the company reports that automated attacks recovering protected chain-of-thought from an earlier request succeeded 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 the updated ones. The change matters beyond OpenAI: the same safeguards now decide what developers can and cannot do with conversation history in the API.

Why reasoning traces are the prize

Distillation means training one model on the outputs of another. Used without permission, it lets a lab copy capabilities without copying the safety rules that came with them. OpenAI’s argument is that the hidden reasoning is the most valuable part, because it contains steps the final answer leaves out.

The company backs this with an internal experiment. It trained the same student model twice on coding and defensive security work from Astra-level teachers: once on what users normally see (answers, reasoning summaries, tool calls) and once on the exact reasoning traces. After fine-tuning, the exact-trace student generally did better on the cyber benchmarks ExploitBench and SEC-bench Pro. After a short round of reinforcement learning the picture was mixed: the exact-trace student led only at the largest inference budget, while the summary-trained one scored higher at lower budgets. OpenAI notes that Anthropic reported a similar advantage for full transcripts on SWE-bench Verified; Anthropic had already described distillation attacks on Claude in February.

The numbers in context: our calculation

OpenAI published two attack types from adaptive black-box red-teaming, aggregated across six attacker models:

  • Reasoning from an earlier request: 54.0% success against GPT-6 Astra, 0.2% against GPT-6.2 Astra. That is roughly a 270-fold drop, from about one attack in two to about one in 500.
  • Reasoning from the current request: 22.9% against 0.9%, about 25 times lower, but still the weaker of the two defenses.

These are OpenAI’s own red-team figures, run by OpenAI’s own automated attackers. No outside lab has reproduced them, and the post does not publish a separate score for GPT-6.2 Sol even though the protections cover it. OpenAI says selected improvements are now also being applied to GPT-6 Astra.

From reports to OpenAI’s own account

When we covered the distillation fight in October, the story rested on press reports that DeepSeek and Moonshot had harvested GPT-6 Astra reasoning through proxy networks, and none of the companies had confirmed anything. The new post fills in OpenAI’s side. It says it attributed a core part of a September extraction campaign to individuals associated with Moonshot AI, and that its weekly enforcement actions against distillation campaigns have ranged from thousands to tens of thousands of accounts, with the rate rising sharply this past summer. It also says it shares indicators of compromise with other model providers. The post gives no further technical detail on that campaign.

What changes for developers using the API

The part most teams will feel is the API-level protection. With GPT-6.2 Astra and Sol, OpenAI ties a model’s protected reasoning to the exact conversation history that produced it, and older model families can no longer read that reasoning. The OpenAI developer guide on prefix changes spells out the consequence: if you shorten a tool result, rewrite a developer message or replace older turns with your own summary, the API drops the reasoning items that came after the change. The model may then redo that thinking, which costs tokens and latency, and output quality can suffer.

In practice that means three checks for agent builders:

  • Compaction: OpenAI’s own compaction endpoint keeps later reasoning; a home-grown summary placed before recent turns does not. The guide suggests putting custom summaries at the end of the context instead.
  • Reasoning effort: switching effort mid-conversation should be done by appending a configuration update, not by changing the request-level setting.
  • Monitoring: the API reports removed reasoning items and the reason, so teams can see when their own history edits are costing them.

What is still open

OpenAI says it plans to share more about recent campaigns it has disrupted. Still unknown: how the 0.2% figure holds up against human attackers rather than automated ones, how often legitimate requests for explanations get caught by the new classifiers, and whether other labs adopt the same conversation-bound approach. For the broader security angle, our Policy & Safety section tracks how labs are hardening their models.

Christopher Ort

Christopher Ort

Christopher Ort is the editor of i10X News at i10x.ai. He has worked hands-on with AI tools, models and agent workflows for years and leads i10X’s coverage of AI launches, funding and research. He is the author of “The Style Penalty”, an SSRN study on how frontier AI models judge AI-written CVs, covered by heise online and t3n.

Timeline

What happened before

  1. Release

    GPT-6.1 Terra

    1.05M context · $1 / $5 per 1M tokens

    openai.com (source, opens in a new tab)
  2. Acquisition

    OpenAI acquires Orchestra Bio

    Acquires Orchestra Bio

    openai.com (source, opens in a new tab)
  3. Release

    GPT-6.1 Sol

    1.05M context · $2 / $10 per 1M tokens

    developers.openai.com (source, opens in a new tab)
  4. Release

    GPT-6 Sol & Luna

    1.05M context · $2 / $10 per 1M tokens

    Two GPT-6 reasoning models. Price shown is GPT-6 Sol (≤272K prompt); GPT-6 Luna costs $0.10 / $0.50.

    developers.openai.com (source, opens in a new tab)
  5. Release

    GPT-6 Astra

    1.05M context · $10 / $50 per 1M tokens

    OpenAI's most capable model for reasoning, coding, computer use and research.

    openai.com (source, opens in a new tab)
Full OpenAI timeline ›

Editor's picks

  1. Georgia Tech’s Academic AI Playbook: 20 Lessons From 67 Real Courses, Free to Reuse
  2. Arcfront’s $1B Run Rate: What the AI Data Startup Confirms and What It Doesn’t
  3. GPT-6.2 Astra Tops Blueprint-Bench 2, 70% Above May Leader GPT-5.5
  4. Deutsche Telekom Puts 100 Million Euros Into Mistral for Sovereign AI in Germany
  5. Inkling 2 Mini: Thinking Machines’ Looped Open Model Tops Qwen on Coding Tests