OpenAI says new safeguards cut reasoning extraction to 0.2%

The company detailed how it fights adversarial distillation, including classifiers that scan agents' tool-call arguments, and said weekly enforcement has reached tens of thousands of accounts.

Illustration of a glowing circuit-patterned brain under a glass dome while small robots aim hoses of light at it from outside. Image: AgenticFeed illustration

OpenAI says updated safeguards on its GPT-6.2 Astra and GPT-6.2 Sol models have cut the success rate of attacks designed to pull out the models' hidden reasoning from 54.0% to 0.2% in internal red-team testing. The company described the defenses in a November 20 post on what it calls adversarial distillation, in which someone trains their own model on a frontier model's outputs without permission.

OpenAI framed the issue as a safety problem, not only an intellectual property one. "Through adversarial distillation, a second model can learn advanced frontier capabilities without inheriting the safeguards that govern how the original model can use them," the company wrote. It pointed to dual-use fields such as cybersecurity and biology, where a copied model could help with tasks the original would refuse.

Why reasoning traces matter

The post puts particular weight on protected reasoning, the intermediate chain-of-thought steps a model produces before it answers. OpenAI said those traces "can contain information the final answer leaves out, making them especially valuable training data for another model."

OpenAI said it fine-tuned the same pretrained student model on trajectories from Astra-level teacher models working on coding and defensive security tasks. After supervised fine-tuning, the student trained on exact reasoning traces generally outperformed one trained on reasoning summaries. After a round of reinforcement learning, the exact-trace student still scored higher at the largest inference budget tested, while the summary-trained student did better at lower budgets, according to the post.

OpenAI cited related Anthropic research that found a similar advantage from full reasoning transcripts on SWE-bench Verified, a software engineering benchmark.

Defenses that cover tool calls

OpenAI said its defenses work at two levels: individual requests and coordinated campaigns. At the request level, adversarial training teaches models to refuse extraction attempts, while input classifiers flag suspicious prompts and output classifiers scan responses for leaked reasoning.

The output checks cover tool-call arguments as well as chat replies. "A model can expose protected reasoning in a tool's arguments without including it in its final answer," the company wrote, a gap that grows in importance as models run as agents that call external tools.

On the API side, OpenAI said GPT-6.2 Astra and Sol now tie protected reasoning to the conversation history that produced it. If an earlier message is rewritten, the API drops the reasoning that followed it. Older model families with weaker safeguards are also blocked from reading the newer models' protected reasoning.

The company said it also closed a pathway that could have exposed another user's encrypted reasoning. That work followed a September disclosure about attacks that moved encrypted reasoning from one conversation into another and asked the model to reveal it.

In adaptive black-box testing aggregated across six attacker models, the success rate for extracting reasoning from the current request fell from 22.9% on GPT-6 Astra to 0.9% on GPT-6.2 Astra, OpenAI reported.

Enforcement and policy

OpenAI said it has seen multiple unauthorized distillation campaigns on its platform this year. Its weekly enforcement actions have ranged from thousands to tens of thousands of accounts, with the rate rising significantly over the summer, according to the post. It did not give a total.

OpenAI also repeated earlier accusations against Chinese developers, saying it briefed Congress in February on activity indicating that DeepSeek was attempting to distill U.S. frontier models, and that in September it attributed a core part of one extraction campaign to individuals associated with Moonshot AI, the developer of Kimi. The post does not include responses from either company, and OpenAI said it plans to share more about recent campaigns soon.

OpenAI also said it supports legislation that would give the U.S. government more tools to identify and sanction foreign actors that illicitly extract capabilities from American models. It said it has told Congress it backs the Chip Security Act among other bills.

For other model developers, the post recommends keeping logs of successful extraction attacks as regression tests and replaying them against updated defenses. It also urges providers to share indicators of compromise with one another.

Source: OpenAI, OpenAI, “Combatting adversarial distillation”

Get the AgenticFeed newsletter

A free, regular roundup of the agent stories that matter. No spam, and you can unsubscribe at any time.

Want to reach readers following AI agents? Advertise with AgenticFeed

Next event · 30 Nov, 2026

AWS re:Invent 2026 in Las Vegas

AWS’s annual conference runs Nov. 30 to Dec. 4. Matt Garman’s CEO keynote is on Dec. 1, and the catalog lists more than 2,200 sessions.

The AgenticFeed newsletter

A free, regular roundup of the agent stories that matter. No spam.