VMTech
Discuss a project →

OpenAI details new safeguards against adversarial AI distillation

OpenAI details new safeguards against adversarial AI distillation

OpenAI has described new protections designed to prevent adversarial distillation of its frontier models, reporting that the average success rate for extracting protected reasoning from an earlier request fell from 54.0% in GPT-6 Astra to 0.2% in GPT-6.2 Astra. In the same adaptive black-box evaluation, extraction of reasoning from the current request fell from 22.9% to 0.9%.

The measures apply to GPT-6.2 Astra and GPT-6.2 Sol and combine model training, classifiers, API controls and investigations of coordinated extraction campaigns. OpenAI said it is also applying selected safeguard improvements to GPT-6 Astra.

Why reasoning traces are a distillation target

Distillation is a standard machine-learning technique in which one model learns from data produced by another. OpenAI’s concern is unauthorized use of a frontier model’s outputs to train a second model with similar capabilities but without the original developer’s safeguards.

The company places particular emphasis on protected reasoning, or the intermediate steps generated while a model works toward an answer. Such traces can contain information omitted from the final response, making them valuable training material for a student model. The risk is particularly significant in dual-use areas including cybersecurity and biology, where capabilities intended for legitimate work could be repurposed if transferred to systems with weaker controls.

Internal cyber capability tests

OpenAI tested capability transfer using trajectories from Astra-level teacher models performing coding and defensive cybersecurity tasks. The training material included tool interactions and final responses, while one student model also received exact reasoning traces and another received reasoning summaries. Both training runs began with the same pretrained student model.

After supervised fine-tuning, the student trained on exact traces generally outperformed the summary-trained student. Following a lightweight reinforcement-learning stage in cyber and coding environments, the exact-trace student remained ahead at the largest inference budget tested on both benchmarks, ExploitBench and SEC-bench Pro. OpenAI said the result suggests that protecting reasoning traces can limit cyber capability transfer even where the original tasks are allowed by its usage policies.

Controls at request and campaign levels

At the request level, OpenAI uses adversarial training to help models recognise extraction attempts and refuse to disclose protected reasoning. Input classifiers look for attempts to obtain it, while output classifiers inspect assistant messages and tool-call arguments for leaks. Tool-call coverage matters because reasoning can appear in arguments without being exposed in a final answer.

API protections also restrict reuse of protected reasoning across interactions and models. For GPT-6.2 Astra and Sol, OpenAI ties protected reasoning to the conversation history that produced it. The model can use earlier reasoning as a conversation develops, but rewriting an earlier message causes the API to drop the reasoning that followed it. Older model families with weaker safeguards are prevented from reading the protected reasoning of the two GPT-6.2 models.

At campaign level, the company investigates coordinated activity using request- and account-level signals, then enforces against networks of accounts when it identifies policy or terms violations. OpenAI said its weekly enforcement actions this year have ranged from thousands to tens of thousands of accounts, with activity increasing significantly during the summer. It also shares indicators of compromise with other model providers where appropriate.

Business implication

Businesses deploying advanced models should treat reasoning-related data, tool arguments and conversation context as sensitive assets: monitor attempts to extract them, enforce access boundaries across users and model versions, and pair automated controls with investigation and response processes.

#aisecurity#modelsecurity#cybersecurity#openai
Open analytics
On the site 2 views
min read 4 20.11.2026
Instagram

OpenAI details new safeguards against adversarial AI distillation

Open the post on Instagram ↗