OpenAI details GPT-6.2 safeguards against model extraction
OpenAI says editing an earlier message can make the model redo work because its API drops the private reasoning that followed that message.
OpenAI on November 20 detailed safeguards it says are intended to limit extraction of protected reasoning from GPT-6.2 Astra and GPT-6.2 Sol. Protected reasoning is the private work a model does to arrive at an answer. OpenAI says that material can be used to train another model to reproduce capabilities without carrying over the original safeguards.
OpenAI says both GPT-6.2 models tie protected reasoning to the conversation history that produced it. The models can build on that reasoning as new messages are added. The company says its API drops the reasoning that followed an earlier message if that message is edited.
OpenAI says this can make the model redo some of its reasoning, which can increase token use, costs or response times. The company also says it blocks older model families with weaker safeguards from reading either GPT-6.2 model’s protected reasoning.
The company says it trains both models to withhold private reasoning while still answering ordinary requests for explanations. OpenAI says it also screens incoming requests and checks model replies and information sent to tools for possible leaks, including during streaming.
OpenAI ran an internal test designed to resemble production in which six automated testers adapted their attempts based on what an outside user could see. OpenAI reported average extraction rates of 0.2% for earlier requests and 0.9% for current requests on GPT-6.2 Astra with updated safeguards. GPT-6 Astra with its original safeguards had rates of 54.0% and 22.9%, respectively.
OpenAI reported rates of 0.5% for earlier requests and 0.8% for current requests on GPT-6.2 Sol with the updated safeguards. Both the model and the safeguards changed in the Astra comparison. The November disclosure gave no extraction rate for live traffic.
The disclosure follows OpenAI’s September 30 account of disrupting a coordinated extraction campaign in July. OpenAI said on November 20 that it was applying some of the safeguard updates to the earlier GPT-6 Astra.
More news

Dragos publishes findings from about 2 million AI test prompts

OpenAI details frontier-model controls from training through deployment

Pydantic becomes a CVE Numbering Authority for its projects
