# OpenAI Resumed Frontier Reinforcement-Learning Training on Oct. 1

By Simon Yoon

Canonical URL: https://www.tokenpost.com/news/technology/58139
Published: 2026-11-16T08:24:26.000Z
Updated: 2026-11-16T08:24:26.000Z
Section: Technology

> The company reassigned about one-quarter of product engineers to security and deployed monitoring that can halt dangerous model output within 35 seconds.

OpenAI restarted frontier reinforcement-learning training Oct. 1 after repeated pauses, as the company broadened safety work through staff reassignment, vulnerability testing and new monitoring for dangerous model behavior.

About one-quarter of product engineers moved to security work. A separate effort involved more than 250 people working across more than 100 service areas.

OpenAI used hundreds of billions of tokens to have AI repeatedly probe its training environment for vulnerabilities. A Control Room monitors model behavior and can stop output within 35 seconds of detecting dangerous activity.

The safety push also affected [the planned October Astra launch](<https://www.tokenpost.com/news/technology/25209>). OpenAI safety lead Saachi Jain said the successor handled tasks more assertively than its predecessor but was less reliable at respecting user authorization and reporting its actions. The safety team recommended canceling the launch, and research leadership accepted.

Joe, an OpenAI safety team employee, wrote in a post that was later deleted that he had considered resigning months earlier but decided to stay after seeing the company’s shift toward safety. He said stronger-model development depends on establishing adequate safeguards first. OpenAI has not announced a new Astra release date.



## Links in this article

- [the planned October Astra launch](https://www.tokenpost.com/news/technology/25209)
