Independent/Reader-funded/Infrastructure, not tokens
DeAINEWS

AI you control — open models, private inference, and the networks that run them.

PulseProvider Policy & Trust

OpenAI details defenses for protected reasoning traces

OpenAI published anti-distillation research: self-reported red-team extraction success fell from 54.0% to 0.2% in GPT-6.2 Astra, with layered trace defenses.

DeAI is powered by Morpheus (mor.org). We cover competing providers on the same terms — see our methodology.

A black mesh security cage in a datacenter aisle enclosing two server racks with amber status lights, one thick cable conduit running from the rack to a sealed wall plate — the secured hardware behind protected reasoning traces. Illustration: DeAI
A black mesh security cage in a datacenter aisle enclosing two server racks with amber status lights, one thick cable conduit running from the rack to a sealed wall plate — the secured hardware behind protected reasoning traces. Illustration: DeAI

OpenAI published anti-distillation research on November 20: its internal red-team reports reasoning-trace extraction success fell from 54.0% to 0.2% in GPT-6.2 Astra. The number is self-reported, but the design change behind it — reasoning traces becoming an API-managed, provider-controlled asset — is verifiable and cuts both ways for builders.

Key facts

What's driving the conversation

The post moved fast — 570K+ views on the main announcement within the sweep window, with @sama amplifying and @swyx among the analysts threading through the implications. Two camps formed. The safety-aligned reading takes the post at face value: distillation of exact traces transfers offensive cyber capability, so hardening traces is defensive work, and OpenAI explicitly cites Anthropic's parallel research as corroboration. The skeptical reading, prominent in the replies, treats the 54.0%→0.2% figure as unfalsifiable marketing for a moat: if your chain-of-thought can't be extracted, it can't be distilled, and closed-model pricing stays closed. Both readings agree on the mechanics; they disagree on the motive.

The substance

What is verifiable is the mechanism. The post describes layered defenses: adversarial training that teaches models to refuse reasoning-extraction attempts, input classifiers that detect extraction prompts and output classifiers that catch leaks in assistant messages and tool-call arguments, and API-level protections that tie a model's protected reasoning to the conversation history that produced it — if you rewrite an earlier message in GPT-6.2, the API drops the reasoning that followed it. OpenAI also blocks older model families from reading GPT-6.2's protected reasoning, and describes campaign-level investigations into coordinated extraction attempts, sharing indicators of compromise with other providers.

What is not verifiable is the headline number. The 54.0%→0.2% figure comes from OpenAI's own adaptive red-team evaluation using its own automated red-teaming stack; the methodology, attack set, and success criteria are not published in a form an outsider can reproduce. Treat it as a claim — plausibly directionally true, but a claim. The internal ExploitBench study has the same status: it is a described experiment, not a released benchmark.

One more framing note: the "this admits CoT is the pricing moat" argument circulating on X is discourse, not OpenAI's position. The paper makes a safety argument throughout — traces as dangerous capability transfer — and never claims pricing motivation. Attributing the moat framing to OpenAI would be wrong; attributing it to the discourse is accurate.

Why builders are watching

The practical consequences run in two directions. If you build on closed APIs, expect reasoning access to keep narrowing: more providers will bind traces to conversations, filter tool-call arguments, and treat extraction as abuse — and if you depend on structured reasoning in tool calls, output classifiers can now block your requests. If you self-host open weights, your traces are yours by construction; the same research is also a roadmap of what trace leakage actually buys an attacker, which open-model deployers should read as a threat model rather than a product announcement. Either way, this extends OpenAI's frontier training security overhaul and sharpens the data-handling questions in whether OpenAI trains on your outputs — with the open-versus-closed capability gap now tracked in our GPT-5.5 comparison as the relevant backdrop for what distillation protection is protecting.

Questions

What did OpenAI announce about adversarial distillation?
On November 20, 2026, OpenAI published defenses protecting its models' reasoning traces from extraction. In an internal adaptive red-team evaluation it reports average extraction success fell from 54.0% against GPT-6 Astra with original safeguards to 0.2% against GPT-6.2 Astra with updated safeguards. The figure is OpenAI's own and not independently reproducible.
Why do reasoning traces matter for distillation?
OpenAI's internal study found students trained on a teacher's exact reasoning traces outperformed students trained on reasoning summaries after reinforcement learning on cyber benchmarks (ExploitBench). Traces contain information the final answer leaves out, so they transfer more capability, including potentially offensive cyber skills.
What is disputed about this research?
The 54.0% to 0.2% improvement is OpenAI's self-reported internal evaluation — the methodology is not independently reproducible. X discourse also framed the work as an admission that chain-of-thought underwrites closed-model pricing moats; the paper itself makes a safety argument and makes no pricing claims.
How does OpenAI technically protect reasoning traces?
Layered defenses: adversarial training against extraction attempts, input and output classifiers covering tool-call arguments, and API-level binding — GPT-6.2 ties protected reasoning to the conversation that produced it, and drops reasoning if an earlier message is rewritten.

Sources

  1. Combatting adversarial distillation — OpenAI
  2. Anthropic distillation defenses (related work) — Anthropic
  3. X discussion of the announcement — X

About DeAI

DeAI is an independent publication covering open-weight AI models, private inference, and decentralized infrastructure — the tools for running AI you actually control. We test providers on price, privacy, and refusal behavior and publish the numbers, not the vibes. DeAI is powered by Morpheus (mor.org), a decentralized inference marketplace, and covers it on the same terms as every other provider.

Powered by Morpheus and StrandCMS

Morpheus is a decentralized inference marketplace, covered on the same terms as every other provider — we rank it wherever the data lands. StrandCMS is the open-source, agent-first framework this site is built on.

Learn more about the Morpheus Inference API →

Sponsor disclosure — not editorial

Powered by Morpheus and StrandCMS. Morpheus is a decentralized inference marketplace, covered on the same terms as every other provider. StrandCMS is the open-source, agent-first framework this site is built on.

Learn more →